1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * NET3 Protocol independent device support routines. 4 * 5 * Derived from the non IP parts of dev.c 1.0.19 6 * Authors: Ross Biro 7 * Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG> 8 * Mark Evans, <evansmp@uhura.aston.ac.uk> 9 * 10 * Additional Authors: 11 * Florian la Roche <rzsfl@rz.uni-sb.de> 12 * Alan Cox <gw4pts@gw4pts.ampr.org> 13 * David Hinds <dahinds@users.sourceforge.net> 14 * Alexey Kuznetsov <kuznet@ms2.inr.ac.ru> 15 * Adam Sulmicki <adam@cfar.umd.edu> 16 * Pekka Riikonen <priikone@poesidon.pspt.fi> 17 * 18 * Changes: 19 * D.J. Barrow : Fixed bug where dev->refcnt gets set 20 * to 2 if register_netdev gets called 21 * before net_dev_init & also removed a 22 * few lines of code in the process. 23 * Alan Cox : device private ioctl copies fields back. 24 * Alan Cox : Transmit queue code does relevant 25 * stunts to keep the queue safe. 26 * Alan Cox : Fixed double lock. 27 * Alan Cox : Fixed promisc NULL pointer trap 28 * ???????? : Support the full private ioctl range 29 * Alan Cox : Moved ioctl permission check into 30 * drivers 31 * Tim Kordas : SIOCADDMULTI/SIOCDELMULTI 32 * Alan Cox : 100 backlog just doesn't cut it when 33 * you start doing multicast video 8) 34 * Alan Cox : Rewrote net_bh and list manager. 35 * Alan Cox : Fix ETH_P_ALL echoback lengths. 36 * Alan Cox : Took out transmit every packet pass 37 * Saved a few bytes in the ioctl handler 38 * Alan Cox : Network driver sets packet type before 39 * calling netif_rx. Saves a function 40 * call a packet. 41 * Alan Cox : Hashed net_bh() 42 * Richard Kooijman: Timestamp fixes. 43 * Alan Cox : Wrong field in SIOCGIFDSTADDR 44 * Alan Cox : Device lock protection. 45 * Alan Cox : Fixed nasty side effect of device close 46 * changes. 47 * Rudi Cilibrasi : Pass the right thing to 48 * set_mac_address() 49 * Dave Miller : 32bit quantity for the device lock to 50 * make it work out on a Sparc. 51 * Bjorn Ekwall : Added KERNELD hack. 52 * Alan Cox : Cleaned up the backlog initialise. 53 * Craig Metz : SIOCGIFCONF fix if space for under 54 * 1 device. 55 * Thomas Bogendoerfer : Return ENODEV for dev_open, if there 56 * is no device open function. 57 * Andi Kleen : Fix error reporting for SIOCGIFCONF 58 * Michael Chastain : Fix signed/unsigned for SIOCGIFCONF 59 * Cyrus Durgin : Cleaned for KMOD 60 * Adam Sulmicki : Bug Fix : Network Device Unload 61 * A network device unload needs to purge 62 * the backlog queue. 63 * Paul Rusty Russell : SIOCSIFNAME 64 * Pekka Riikonen : Netdev boot-time settings code 65 * Andrew Morton : Make unregister_netdevice wait 66 * indefinitely on dev->refcnt 67 * J Hadi Salim : - Backlog queue sampling 68 * - netif_rx() feedback 69 */ 70 71 #include <linux/uaccess.h> 72 #include <linux/bitmap.h> 73 #include <linux/capability.h> 74 #include <linux/cpu.h> 75 #include <linux/types.h> 76 #include <linux/kernel.h> 77 #include <linux/hash.h> 78 #include <linux/slab.h> 79 #include <linux/sched.h> 80 #include <linux/sched/isolation.h> 81 #include <linux/sched/mm.h> 82 #include <linux/smpboot.h> 83 #include <linux/mutex.h> 84 #include <linux/rwsem.h> 85 #include <linux/string.h> 86 #include <linux/mm.h> 87 #include <linux/socket.h> 88 #include <linux/sockios.h> 89 #include <linux/errno.h> 90 #include <linux/interrupt.h> 91 #include <linux/if_ether.h> 92 #include <linux/netdevice.h> 93 #include <linux/etherdevice.h> 94 #include <linux/ethtool.h> 95 #include <linux/ethtool_netlink.h> 96 #include <linux/skbuff.h> 97 #include <linux/kthread.h> 98 #include <linux/bpf.h> 99 #include <linux/bpf_trace.h> 100 #include <net/net_namespace.h> 101 #include <net/sock.h> 102 #include <net/busy_poll.h> 103 #include <linux/rtnetlink.h> 104 #include <linux/stat.h> 105 #include <net/dsa.h> 106 #include <net/dst.h> 107 #include <net/dst_metadata.h> 108 #include <net/gro.h> 109 #include <net/netdev_queues.h> 110 #include <net/pkt_sched.h> 111 #include <net/pkt_cls.h> 112 #include <net/checksum.h> 113 #include <net/xfrm.h> 114 #include <net/tcx.h> 115 #include <linux/highmem.h> 116 #include <linux/init.h> 117 #include <linux/module.h> 118 #include <linux/netpoll.h> 119 #include <linux/rcupdate.h> 120 #include <linux/delay.h> 121 #include <net/iw_handler.h> 122 #include <asm/current.h> 123 #include <linux/audit.h> 124 #include <linux/dmaengine.h> 125 #include <linux/err.h> 126 #include <linux/ctype.h> 127 #include <linux/if_arp.h> 128 #include <linux/if_vlan.h> 129 #include <linux/ip.h> 130 #include <net/ip.h> 131 #include <net/mpls.h> 132 #include <linux/ipv6.h> 133 #include <linux/in.h> 134 #include <linux/jhash.h> 135 #include <linux/random.h> 136 #include <trace/events/napi.h> 137 #include <trace/events/net.h> 138 #include <trace/events/skb.h> 139 #include <trace/events/qdisc.h> 140 #include <trace/events/xdp.h> 141 #include <linux/inetdevice.h> 142 #include <linux/cpu_rmap.h> 143 #include <linux/static_key.h> 144 #include <linux/hashtable.h> 145 #include <linux/vmalloc.h> 146 #include <linux/if_macvlan.h> 147 #include <linux/errqueue.h> 148 #include <linux/hrtimer.h> 149 #include <linux/netfilter_netdev.h> 150 #include <linux/crash_dump.h> 151 #include <linux/sctp.h> 152 #include <net/udp_tunnel.h> 153 #include <linux/net_namespace.h> 154 #include <linux/indirect_call_wrapper.h> 155 #include <net/devlink.h> 156 #include <linux/pm_runtime.h> 157 #include <linux/prandom.h> 158 #include <linux/once_lite.h> 159 #include <net/netdev_lock.h> 160 #include <net/netdev_rx_queue.h> 161 #include <net/page_pool/types.h> 162 #include <net/page_pool/helpers.h> 163 #include <net/page_pool/memory_provider.h> 164 #include <net/rps.h> 165 #include <linux/phy_link_topology.h> 166 167 #include "dev.h" 168 #include "devmem.h" 169 #include "net-sysfs.h" 170 171 static DEFINE_SPINLOCK(ptype_lock); 172 struct list_head ptype_base[PTYPE_HASH_SIZE] __read_mostly; 173 174 static int netif_rx_internal(struct sk_buff *skb); 175 static int call_netdevice_notifiers_extack(unsigned long val, 176 struct net_device *dev, 177 struct netlink_ext_ack *extack); 178 179 static DEFINE_MUTEX(ifalias_mutex); 180 181 /* protects napi_hash addition/deletion and napi_gen_id */ 182 static DEFINE_SPINLOCK(napi_hash_lock); 183 184 static unsigned int napi_gen_id = NR_CPUS; 185 static DEFINE_READ_MOSTLY_HASHTABLE(napi_hash, 8); 186 187 static inline void dev_base_seq_inc(struct net *net) 188 { 189 unsigned int val = net->dev_base_seq + 1; 190 191 WRITE_ONCE(net->dev_base_seq, val ?: 1); 192 } 193 194 static inline struct hlist_head *dev_name_hash(struct net *net, const char *name) 195 { 196 unsigned int hash = full_name_hash(net, name, strnlen(name, IFNAMSIZ)); 197 198 return &net->dev_name_head[hash_32(hash, NETDEV_HASHBITS)]; 199 } 200 201 static inline struct hlist_head *dev_index_hash(struct net *net, int ifindex) 202 { 203 return &net->dev_index_head[ifindex & (NETDEV_HASHENTRIES - 1)]; 204 } 205 206 #ifndef CONFIG_PREEMPT_RT 207 208 static DEFINE_STATIC_KEY_FALSE(use_backlog_threads_key); 209 210 static int __init setup_backlog_napi_threads(char *arg) 211 { 212 static_branch_enable(&use_backlog_threads_key); 213 return 0; 214 } 215 early_param("thread_backlog_napi", setup_backlog_napi_threads); 216 217 static bool use_backlog_threads(void) 218 { 219 return static_branch_unlikely(&use_backlog_threads_key); 220 } 221 222 #else 223 224 static bool use_backlog_threads(void) 225 { 226 return true; 227 } 228 229 #endif 230 231 static inline void backlog_lock_irq_save(struct softnet_data *sd, 232 unsigned long *flags) 233 { 234 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 235 spin_lock_irqsave(&sd->input_pkt_queue.lock, *flags); 236 else 237 local_irq_save(*flags); 238 } 239 240 static inline void backlog_lock_irq_disable(struct softnet_data *sd) 241 { 242 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 243 spin_lock_irq(&sd->input_pkt_queue.lock); 244 else 245 local_irq_disable(); 246 } 247 248 static inline void backlog_unlock_irq_restore(struct softnet_data *sd, 249 unsigned long *flags) 250 { 251 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 252 spin_unlock_irqrestore(&sd->input_pkt_queue.lock, *flags); 253 else 254 local_irq_restore(*flags); 255 } 256 257 static inline void backlog_unlock_irq_enable(struct softnet_data *sd) 258 { 259 if (IS_ENABLED(CONFIG_RPS) || use_backlog_threads()) 260 spin_unlock_irq(&sd->input_pkt_queue.lock); 261 else 262 local_irq_enable(); 263 } 264 265 static struct netdev_name_node *netdev_name_node_alloc(struct net_device *dev, 266 const char *name) 267 { 268 struct netdev_name_node *name_node; 269 270 name_node = kmalloc(sizeof(*name_node), GFP_KERNEL); 271 if (!name_node) 272 return NULL; 273 INIT_HLIST_NODE(&name_node->hlist); 274 name_node->dev = dev; 275 name_node->name = name; 276 return name_node; 277 } 278 279 static struct netdev_name_node * 280 netdev_name_node_head_alloc(struct net_device *dev) 281 { 282 struct netdev_name_node *name_node; 283 284 name_node = netdev_name_node_alloc(dev, dev->name); 285 if (!name_node) 286 return NULL; 287 INIT_LIST_HEAD(&name_node->list); 288 return name_node; 289 } 290 291 static void netdev_name_node_free(struct netdev_name_node *name_node) 292 { 293 kfree(name_node); 294 } 295 296 static void netdev_name_node_add(struct net *net, 297 struct netdev_name_node *name_node) 298 { 299 hlist_add_head_rcu(&name_node->hlist, 300 dev_name_hash(net, name_node->name)); 301 } 302 303 static void netdev_name_node_del(struct netdev_name_node *name_node) 304 { 305 hlist_del_rcu(&name_node->hlist); 306 } 307 308 static struct netdev_name_node *netdev_name_node_lookup(struct net *net, 309 const char *name) 310 { 311 struct hlist_head *head = dev_name_hash(net, name); 312 struct netdev_name_node *name_node; 313 314 hlist_for_each_entry(name_node, head, hlist) 315 if (!strcmp(name_node->name, name)) 316 return name_node; 317 return NULL; 318 } 319 320 static struct netdev_name_node *netdev_name_node_lookup_rcu(struct net *net, 321 const char *name) 322 { 323 struct hlist_head *head = dev_name_hash(net, name); 324 struct netdev_name_node *name_node; 325 326 hlist_for_each_entry_rcu(name_node, head, hlist) 327 if (!strcmp(name_node->name, name)) 328 return name_node; 329 return NULL; 330 } 331 332 bool netdev_name_in_use(struct net *net, const char *name) 333 { 334 return netdev_name_node_lookup(net, name); 335 } 336 EXPORT_SYMBOL(netdev_name_in_use); 337 338 int netdev_name_node_alt_create(struct net_device *dev, const char *name) 339 { 340 struct netdev_name_node *name_node; 341 struct net *net = dev_net(dev); 342 343 name_node = netdev_name_node_lookup(net, name); 344 if (name_node) 345 return -EEXIST; 346 name_node = netdev_name_node_alloc(dev, name); 347 if (!name_node) 348 return -ENOMEM; 349 netdev_name_node_add(net, name_node); 350 /* The node that holds dev->name acts as a head of per-device list. */ 351 list_add_tail_rcu(&name_node->list, &dev->name_node->list); 352 353 return 0; 354 } 355 356 static void netdev_name_node_alt_free(struct rcu_head *head) 357 { 358 struct netdev_name_node *name_node = 359 container_of(head, struct netdev_name_node, rcu); 360 361 kfree(name_node->name); 362 netdev_name_node_free(name_node); 363 } 364 365 static void __netdev_name_node_alt_destroy(struct netdev_name_node *name_node) 366 { 367 netdev_name_node_del(name_node); 368 list_del(&name_node->list); 369 call_rcu(&name_node->rcu, netdev_name_node_alt_free); 370 } 371 372 int netdev_name_node_alt_destroy(struct net_device *dev, const char *name) 373 { 374 struct netdev_name_node *name_node; 375 struct net *net = dev_net(dev); 376 377 name_node = netdev_name_node_lookup(net, name); 378 if (!name_node) 379 return -ENOENT; 380 /* lookup might have found our primary name or a name belonging 381 * to another device. 382 */ 383 if (name_node == dev->name_node || name_node->dev != dev) 384 return -EINVAL; 385 386 __netdev_name_node_alt_destroy(name_node); 387 return 0; 388 } 389 390 static void netdev_name_node_alt_flush(struct net_device *dev) 391 { 392 struct netdev_name_node *name_node, *tmp; 393 394 list_for_each_entry_safe(name_node, tmp, &dev->name_node->list, list) { 395 list_del(&name_node->list); 396 netdev_name_node_alt_free(&name_node->rcu); 397 } 398 } 399 400 /* Device list insertion */ 401 static void list_netdevice(struct net_device *dev) 402 { 403 struct netdev_name_node *name_node; 404 struct net *net = dev_net(dev); 405 406 ASSERT_RTNL(); 407 408 list_add_tail_rcu(&dev->dev_list, &net->dev_base_head); 409 netdev_name_node_add(net, dev->name_node); 410 hlist_add_head_rcu(&dev->index_hlist, 411 dev_index_hash(net, dev->ifindex)); 412 413 netdev_for_each_altname(dev, name_node) 414 netdev_name_node_add(net, name_node); 415 416 /* We reserved the ifindex, this can't fail */ 417 WARN_ON(xa_store(&net->dev_by_index, dev->ifindex, dev, GFP_KERNEL)); 418 419 dev_base_seq_inc(net); 420 } 421 422 /* Device list removal 423 * caller must respect a RCU grace period before freeing/reusing dev 424 */ 425 static void unlist_netdevice(struct net_device *dev) 426 { 427 struct netdev_name_node *name_node; 428 struct net *net = dev_net(dev); 429 430 ASSERT_RTNL(); 431 432 xa_erase(&net->dev_by_index, dev->ifindex); 433 434 netdev_for_each_altname(dev, name_node) 435 netdev_name_node_del(name_node); 436 437 /* Unlink dev from the device chain */ 438 list_del_rcu(&dev->dev_list); 439 netdev_name_node_del(dev->name_node); 440 hlist_del_rcu(&dev->index_hlist); 441 442 dev_base_seq_inc(dev_net(dev)); 443 } 444 445 /* 446 * Our notifier list 447 */ 448 449 static RAW_NOTIFIER_HEAD(netdev_chain); 450 451 /* 452 * Device drivers call our routines to queue packets here. We empty the 453 * queue in the local softnet handler. 454 */ 455 456 DEFINE_PER_CPU_ALIGNED(struct softnet_data, softnet_data) = { 457 .process_queue_bh_lock = INIT_LOCAL_LOCK(process_queue_bh_lock), 458 }; 459 EXPORT_PER_CPU_SYMBOL(softnet_data); 460 461 /* Page_pool has a lockless array/stack to alloc/recycle pages. 462 * PP consumers must pay attention to run APIs in the appropriate context 463 * (e.g. NAPI context). 464 */ 465 DEFINE_PER_CPU(struct page_pool *, system_page_pool); 466 467 #ifdef CONFIG_LOCKDEP 468 /* 469 * register_netdevice() inits txq->_xmit_lock and sets lockdep class 470 * according to dev->type 471 */ 472 static const unsigned short netdev_lock_type[] = { 473 ARPHRD_NETROM, ARPHRD_ETHER, ARPHRD_EETHER, ARPHRD_AX25, 474 ARPHRD_PRONET, ARPHRD_CHAOS, ARPHRD_IEEE802, ARPHRD_ARCNET, 475 ARPHRD_APPLETLK, ARPHRD_DLCI, ARPHRD_ATM, ARPHRD_METRICOM, 476 ARPHRD_IEEE1394, ARPHRD_EUI64, ARPHRD_INFINIBAND, ARPHRD_SLIP, 477 ARPHRD_CSLIP, ARPHRD_SLIP6, ARPHRD_CSLIP6, ARPHRD_RSRVD, 478 ARPHRD_ADAPT, ARPHRD_ROSE, ARPHRD_X25, ARPHRD_HWX25, 479 ARPHRD_PPP, ARPHRD_CISCO, ARPHRD_LAPB, ARPHRD_DDCMP, 480 ARPHRD_RAWHDLC, ARPHRD_TUNNEL, ARPHRD_TUNNEL6, ARPHRD_FRAD, 481 ARPHRD_SKIP, ARPHRD_LOOPBACK, ARPHRD_LOCALTLK, ARPHRD_FDDI, 482 ARPHRD_BIF, ARPHRD_SIT, ARPHRD_IPDDP, ARPHRD_IPGRE, 483 ARPHRD_PIMREG, ARPHRD_HIPPI, ARPHRD_ASH, ARPHRD_ECONET, 484 ARPHRD_IRDA, ARPHRD_FCPP, ARPHRD_FCAL, ARPHRD_FCPL, 485 ARPHRD_FCFABRIC, ARPHRD_IEEE80211, ARPHRD_IEEE80211_PRISM, 486 ARPHRD_IEEE80211_RADIOTAP, ARPHRD_PHONET, ARPHRD_PHONET_PIPE, 487 ARPHRD_IEEE802154, ARPHRD_VOID, ARPHRD_NONE}; 488 489 static const char *const netdev_lock_name[] = { 490 "_xmit_NETROM", "_xmit_ETHER", "_xmit_EETHER", "_xmit_AX25", 491 "_xmit_PRONET", "_xmit_CHAOS", "_xmit_IEEE802", "_xmit_ARCNET", 492 "_xmit_APPLETLK", "_xmit_DLCI", "_xmit_ATM", "_xmit_METRICOM", 493 "_xmit_IEEE1394", "_xmit_EUI64", "_xmit_INFINIBAND", "_xmit_SLIP", 494 "_xmit_CSLIP", "_xmit_SLIP6", "_xmit_CSLIP6", "_xmit_RSRVD", 495 "_xmit_ADAPT", "_xmit_ROSE", "_xmit_X25", "_xmit_HWX25", 496 "_xmit_PPP", "_xmit_CISCO", "_xmit_LAPB", "_xmit_DDCMP", 497 "_xmit_RAWHDLC", "_xmit_TUNNEL", "_xmit_TUNNEL6", "_xmit_FRAD", 498 "_xmit_SKIP", "_xmit_LOOPBACK", "_xmit_LOCALTLK", "_xmit_FDDI", 499 "_xmit_BIF", "_xmit_SIT", "_xmit_IPDDP", "_xmit_IPGRE", 500 "_xmit_PIMREG", "_xmit_HIPPI", "_xmit_ASH", "_xmit_ECONET", 501 "_xmit_IRDA", "_xmit_FCPP", "_xmit_FCAL", "_xmit_FCPL", 502 "_xmit_FCFABRIC", "_xmit_IEEE80211", "_xmit_IEEE80211_PRISM", 503 "_xmit_IEEE80211_RADIOTAP", "_xmit_PHONET", "_xmit_PHONET_PIPE", 504 "_xmit_IEEE802154", "_xmit_VOID", "_xmit_NONE"}; 505 506 static struct lock_class_key netdev_xmit_lock_key[ARRAY_SIZE(netdev_lock_type)]; 507 static struct lock_class_key netdev_addr_lock_key[ARRAY_SIZE(netdev_lock_type)]; 508 509 static inline unsigned short netdev_lock_pos(unsigned short dev_type) 510 { 511 int i; 512 513 for (i = 0; i < ARRAY_SIZE(netdev_lock_type); i++) 514 if (netdev_lock_type[i] == dev_type) 515 return i; 516 /* the last key is used by default */ 517 return ARRAY_SIZE(netdev_lock_type) - 1; 518 } 519 520 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock, 521 unsigned short dev_type) 522 { 523 int i; 524 525 i = netdev_lock_pos(dev_type); 526 lockdep_set_class_and_name(lock, &netdev_xmit_lock_key[i], 527 netdev_lock_name[i]); 528 } 529 530 static inline void netdev_set_addr_lockdep_class(struct net_device *dev) 531 { 532 int i; 533 534 i = netdev_lock_pos(dev->type); 535 lockdep_set_class_and_name(&dev->addr_list_lock, 536 &netdev_addr_lock_key[i], 537 netdev_lock_name[i]); 538 } 539 #else 540 static inline void netdev_set_xmit_lockdep_class(spinlock_t *lock, 541 unsigned short dev_type) 542 { 543 } 544 545 static inline void netdev_set_addr_lockdep_class(struct net_device *dev) 546 { 547 } 548 #endif 549 550 /******************************************************************************* 551 * 552 * Protocol management and registration routines 553 * 554 *******************************************************************************/ 555 556 557 /* 558 * Add a protocol ID to the list. Now that the input handler is 559 * smarter we can dispense with all the messy stuff that used to be 560 * here. 561 * 562 * BEWARE!!! Protocol handlers, mangling input packets, 563 * MUST BE last in hash buckets and checking protocol handlers 564 * MUST start from promiscuous ptype_all chain in net_bh. 565 * It is true now, do not change it. 566 * Explanation follows: if protocol handler, mangling packet, will 567 * be the first on list, it is not able to sense, that packet 568 * is cloned and should be copied-on-write, so that it will 569 * change it and subsequent readers will get broken packet. 570 * --ANK (980803) 571 */ 572 573 static inline struct list_head *ptype_head(const struct packet_type *pt) 574 { 575 if (pt->type == htons(ETH_P_ALL)) 576 return pt->dev ? &pt->dev->ptype_all : &net_hotdata.ptype_all; 577 else 578 return pt->dev ? &pt->dev->ptype_specific : 579 &ptype_base[ntohs(pt->type) & PTYPE_HASH_MASK]; 580 } 581 582 /** 583 * dev_add_pack - add packet handler 584 * @pt: packet type declaration 585 * 586 * Add a protocol handler to the networking stack. The passed &packet_type 587 * is linked into kernel lists and may not be freed until it has been 588 * removed from the kernel lists. 589 * 590 * This call does not sleep therefore it can not 591 * guarantee all CPU's that are in middle of receiving packets 592 * will see the new packet type (until the next received packet). 593 */ 594 595 void dev_add_pack(struct packet_type *pt) 596 { 597 struct list_head *head = ptype_head(pt); 598 599 spin_lock(&ptype_lock); 600 list_add_rcu(&pt->list, head); 601 spin_unlock(&ptype_lock); 602 } 603 EXPORT_SYMBOL(dev_add_pack); 604 605 /** 606 * __dev_remove_pack - remove packet handler 607 * @pt: packet type declaration 608 * 609 * Remove a protocol handler that was previously added to the kernel 610 * protocol handlers by dev_add_pack(). The passed &packet_type is removed 611 * from the kernel lists and can be freed or reused once this function 612 * returns. 613 * 614 * The packet type might still be in use by receivers 615 * and must not be freed until after all the CPU's have gone 616 * through a quiescent state. 617 */ 618 void __dev_remove_pack(struct packet_type *pt) 619 { 620 struct list_head *head = ptype_head(pt); 621 struct packet_type *pt1; 622 623 spin_lock(&ptype_lock); 624 625 list_for_each_entry(pt1, head, list) { 626 if (pt == pt1) { 627 list_del_rcu(&pt->list); 628 goto out; 629 } 630 } 631 632 pr_warn("dev_remove_pack: %p not found\n", pt); 633 out: 634 spin_unlock(&ptype_lock); 635 } 636 EXPORT_SYMBOL(__dev_remove_pack); 637 638 /** 639 * dev_remove_pack - remove packet handler 640 * @pt: packet type declaration 641 * 642 * Remove a protocol handler that was previously added to the kernel 643 * protocol handlers by dev_add_pack(). The passed &packet_type is removed 644 * from the kernel lists and can be freed or reused once this function 645 * returns. 646 * 647 * This call sleeps to guarantee that no CPU is looking at the packet 648 * type after return. 649 */ 650 void dev_remove_pack(struct packet_type *pt) 651 { 652 __dev_remove_pack(pt); 653 654 synchronize_net(); 655 } 656 EXPORT_SYMBOL(dev_remove_pack); 657 658 659 /******************************************************************************* 660 * 661 * Device Interface Subroutines 662 * 663 *******************************************************************************/ 664 665 /** 666 * dev_get_iflink - get 'iflink' value of a interface 667 * @dev: targeted interface 668 * 669 * Indicates the ifindex the interface is linked to. 670 * Physical interfaces have the same 'ifindex' and 'iflink' values. 671 */ 672 673 int dev_get_iflink(const struct net_device *dev) 674 { 675 if (dev->netdev_ops && dev->netdev_ops->ndo_get_iflink) 676 return dev->netdev_ops->ndo_get_iflink(dev); 677 678 return READ_ONCE(dev->ifindex); 679 } 680 EXPORT_SYMBOL(dev_get_iflink); 681 682 /** 683 * dev_fill_metadata_dst - Retrieve tunnel egress information. 684 * @dev: targeted interface 685 * @skb: The packet. 686 * 687 * For better visibility of tunnel traffic OVS needs to retrieve 688 * egress tunnel information for a packet. Following API allows 689 * user to get this info. 690 */ 691 int dev_fill_metadata_dst(struct net_device *dev, struct sk_buff *skb) 692 { 693 struct ip_tunnel_info *info; 694 695 if (!dev->netdev_ops || !dev->netdev_ops->ndo_fill_metadata_dst) 696 return -EINVAL; 697 698 info = skb_tunnel_info_unclone(skb); 699 if (!info) 700 return -ENOMEM; 701 if (unlikely(!(info->mode & IP_TUNNEL_INFO_TX))) 702 return -EINVAL; 703 704 return dev->netdev_ops->ndo_fill_metadata_dst(dev, skb); 705 } 706 EXPORT_SYMBOL_GPL(dev_fill_metadata_dst); 707 708 static struct net_device_path *dev_fwd_path(struct net_device_path_stack *stack) 709 { 710 int k = stack->num_paths++; 711 712 if (WARN_ON_ONCE(k >= NET_DEVICE_PATH_STACK_MAX)) 713 return NULL; 714 715 return &stack->path[k]; 716 } 717 718 int dev_fill_forward_path(const struct net_device *dev, const u8 *daddr, 719 struct net_device_path_stack *stack) 720 { 721 const struct net_device *last_dev; 722 struct net_device_path_ctx ctx = { 723 .dev = dev, 724 }; 725 struct net_device_path *path; 726 int ret = 0; 727 728 memcpy(ctx.daddr, daddr, sizeof(ctx.daddr)); 729 stack->num_paths = 0; 730 while (ctx.dev && ctx.dev->netdev_ops->ndo_fill_forward_path) { 731 last_dev = ctx.dev; 732 path = dev_fwd_path(stack); 733 if (!path) 734 return -1; 735 736 memset(path, 0, sizeof(struct net_device_path)); 737 ret = ctx.dev->netdev_ops->ndo_fill_forward_path(&ctx, path); 738 if (ret < 0) 739 return -1; 740 741 if (WARN_ON_ONCE(last_dev == ctx.dev)) 742 return -1; 743 } 744 745 if (!ctx.dev) 746 return ret; 747 748 path = dev_fwd_path(stack); 749 if (!path) 750 return -1; 751 path->type = DEV_PATH_ETHERNET; 752 path->dev = ctx.dev; 753 754 return ret; 755 } 756 EXPORT_SYMBOL_GPL(dev_fill_forward_path); 757 758 /* must be called under rcu_read_lock(), as we dont take a reference */ 759 static struct napi_struct *napi_by_id(unsigned int napi_id) 760 { 761 unsigned int hash = napi_id % HASH_SIZE(napi_hash); 762 struct napi_struct *napi; 763 764 hlist_for_each_entry_rcu(napi, &napi_hash[hash], napi_hash_node) 765 if (napi->napi_id == napi_id) 766 return napi; 767 768 return NULL; 769 } 770 771 /* must be called under rcu_read_lock(), as we dont take a reference */ 772 static struct napi_struct * 773 netdev_napi_by_id(struct net *net, unsigned int napi_id) 774 { 775 struct napi_struct *napi; 776 777 napi = napi_by_id(napi_id); 778 if (!napi) 779 return NULL; 780 781 if (WARN_ON_ONCE(!napi->dev)) 782 return NULL; 783 if (!net_eq(net, dev_net(napi->dev))) 784 return NULL; 785 786 return napi; 787 } 788 789 /** 790 * netdev_napi_by_id_lock() - find a device by NAPI ID and lock it 791 * @net: the applicable net namespace 792 * @napi_id: ID of a NAPI of a target device 793 * 794 * Find a NAPI instance with @napi_id. Lock its device. 795 * The device must be in %NETREG_REGISTERED state for lookup to succeed. 796 * netdev_unlock() must be called to release it. 797 * 798 * Return: pointer to NAPI, its device with lock held, NULL if not found. 799 */ 800 struct napi_struct * 801 netdev_napi_by_id_lock(struct net *net, unsigned int napi_id) 802 { 803 struct napi_struct *napi; 804 struct net_device *dev; 805 806 rcu_read_lock(); 807 napi = netdev_napi_by_id(net, napi_id); 808 if (!napi || READ_ONCE(napi->dev->reg_state) != NETREG_REGISTERED) { 809 rcu_read_unlock(); 810 return NULL; 811 } 812 813 dev = napi->dev; 814 dev_hold(dev); 815 rcu_read_unlock(); 816 817 dev = __netdev_put_lock(dev); 818 if (!dev) 819 return NULL; 820 821 rcu_read_lock(); 822 napi = netdev_napi_by_id(net, napi_id); 823 if (napi && napi->dev != dev) 824 napi = NULL; 825 rcu_read_unlock(); 826 827 if (!napi) 828 netdev_unlock(dev); 829 return napi; 830 } 831 832 /** 833 * __dev_get_by_name - find a device by its name 834 * @net: the applicable net namespace 835 * @name: name to find 836 * 837 * Find an interface by name. Must be called under RTNL semaphore. 838 * If the name is found a pointer to the device is returned. 839 * If the name is not found then %NULL is returned. The 840 * reference counters are not incremented so the caller must be 841 * careful with locks. 842 */ 843 844 struct net_device *__dev_get_by_name(struct net *net, const char *name) 845 { 846 struct netdev_name_node *node_name; 847 848 node_name = netdev_name_node_lookup(net, name); 849 return node_name ? node_name->dev : NULL; 850 } 851 EXPORT_SYMBOL(__dev_get_by_name); 852 853 /** 854 * dev_get_by_name_rcu - find a device by its name 855 * @net: the applicable net namespace 856 * @name: name to find 857 * 858 * Find an interface by name. 859 * If the name is found a pointer to the device is returned. 860 * If the name is not found then %NULL is returned. 861 * The reference counters are not incremented so the caller must be 862 * careful with locks. The caller must hold RCU lock. 863 */ 864 865 struct net_device *dev_get_by_name_rcu(struct net *net, const char *name) 866 { 867 struct netdev_name_node *node_name; 868 869 node_name = netdev_name_node_lookup_rcu(net, name); 870 return node_name ? node_name->dev : NULL; 871 } 872 EXPORT_SYMBOL(dev_get_by_name_rcu); 873 874 /* Deprecated for new users, call netdev_get_by_name() instead */ 875 struct net_device *dev_get_by_name(struct net *net, const char *name) 876 { 877 struct net_device *dev; 878 879 rcu_read_lock(); 880 dev = dev_get_by_name_rcu(net, name); 881 dev_hold(dev); 882 rcu_read_unlock(); 883 return dev; 884 } 885 EXPORT_SYMBOL(dev_get_by_name); 886 887 /** 888 * netdev_get_by_name() - find a device by its name 889 * @net: the applicable net namespace 890 * @name: name to find 891 * @tracker: tracking object for the acquired reference 892 * @gfp: allocation flags for the tracker 893 * 894 * Find an interface by name. This can be called from any 895 * context and does its own locking. The returned handle has 896 * the usage count incremented and the caller must use netdev_put() to 897 * release it when it is no longer needed. %NULL is returned if no 898 * matching device is found. 899 */ 900 struct net_device *netdev_get_by_name(struct net *net, const char *name, 901 netdevice_tracker *tracker, gfp_t gfp) 902 { 903 struct net_device *dev; 904 905 dev = dev_get_by_name(net, name); 906 if (dev) 907 netdev_tracker_alloc(dev, tracker, gfp); 908 return dev; 909 } 910 EXPORT_SYMBOL(netdev_get_by_name); 911 912 /** 913 * __dev_get_by_index - find a device by its ifindex 914 * @net: the applicable net namespace 915 * @ifindex: index of device 916 * 917 * Search for an interface by index. Returns %NULL if the device 918 * is not found or a pointer to the device. The device has not 919 * had its reference counter increased so the caller must be careful 920 * about locking. The caller must hold the RTNL semaphore. 921 */ 922 923 struct net_device *__dev_get_by_index(struct net *net, int ifindex) 924 { 925 struct net_device *dev; 926 struct hlist_head *head = dev_index_hash(net, ifindex); 927 928 hlist_for_each_entry(dev, head, index_hlist) 929 if (dev->ifindex == ifindex) 930 return dev; 931 932 return NULL; 933 } 934 EXPORT_SYMBOL(__dev_get_by_index); 935 936 /** 937 * dev_get_by_index_rcu - find a device by its ifindex 938 * @net: the applicable net namespace 939 * @ifindex: index of device 940 * 941 * Search for an interface by index. Returns %NULL if the device 942 * is not found or a pointer to the device. The device has not 943 * had its reference counter increased so the caller must be careful 944 * about locking. The caller must hold RCU lock. 945 */ 946 947 struct net_device *dev_get_by_index_rcu(struct net *net, int ifindex) 948 { 949 struct net_device *dev; 950 struct hlist_head *head = dev_index_hash(net, ifindex); 951 952 hlist_for_each_entry_rcu(dev, head, index_hlist) 953 if (dev->ifindex == ifindex) 954 return dev; 955 956 return NULL; 957 } 958 EXPORT_SYMBOL(dev_get_by_index_rcu); 959 960 /* Deprecated for new users, call netdev_get_by_index() instead */ 961 struct net_device *dev_get_by_index(struct net *net, int ifindex) 962 { 963 struct net_device *dev; 964 965 rcu_read_lock(); 966 dev = dev_get_by_index_rcu(net, ifindex); 967 dev_hold(dev); 968 rcu_read_unlock(); 969 return dev; 970 } 971 EXPORT_SYMBOL(dev_get_by_index); 972 973 /** 974 * netdev_get_by_index() - find a device by its ifindex 975 * @net: the applicable net namespace 976 * @ifindex: index of device 977 * @tracker: tracking object for the acquired reference 978 * @gfp: allocation flags for the tracker 979 * 980 * Search for an interface by index. Returns NULL if the device 981 * is not found or a pointer to the device. The device returned has 982 * had a reference added and the pointer is safe until the user calls 983 * netdev_put() to indicate they have finished with it. 984 */ 985 struct net_device *netdev_get_by_index(struct net *net, int ifindex, 986 netdevice_tracker *tracker, gfp_t gfp) 987 { 988 struct net_device *dev; 989 990 dev = dev_get_by_index(net, ifindex); 991 if (dev) 992 netdev_tracker_alloc(dev, tracker, gfp); 993 return dev; 994 } 995 EXPORT_SYMBOL(netdev_get_by_index); 996 997 /** 998 * dev_get_by_napi_id - find a device by napi_id 999 * @napi_id: ID of the NAPI struct 1000 * 1001 * Search for an interface by NAPI ID. Returns %NULL if the device 1002 * is not found or a pointer to the device. The device has not had 1003 * its reference counter increased so the caller must be careful 1004 * about locking. The caller must hold RCU lock. 1005 */ 1006 struct net_device *dev_get_by_napi_id(unsigned int napi_id) 1007 { 1008 struct napi_struct *napi; 1009 1010 WARN_ON_ONCE(!rcu_read_lock_held()); 1011 1012 if (!napi_id_valid(napi_id)) 1013 return NULL; 1014 1015 napi = napi_by_id(napi_id); 1016 1017 return napi ? napi->dev : NULL; 1018 } 1019 1020 /* Release the held reference on the net_device, and if the net_device 1021 * is still registered try to lock the instance lock. If device is being 1022 * unregistered NULL will be returned (but the reference has been released, 1023 * either way!) 1024 * 1025 * This helper is intended for locking net_device after it has been looked up 1026 * using a lockless lookup helper. Lock prevents the instance from going away. 1027 */ 1028 struct net_device *__netdev_put_lock(struct net_device *dev) 1029 { 1030 netdev_lock(dev); 1031 if (dev->reg_state > NETREG_REGISTERED) { 1032 netdev_unlock(dev); 1033 dev_put(dev); 1034 return NULL; 1035 } 1036 dev_put(dev); 1037 return dev; 1038 } 1039 1040 /** 1041 * netdev_get_by_index_lock() - find a device by its ifindex 1042 * @net: the applicable net namespace 1043 * @ifindex: index of device 1044 * 1045 * Search for an interface by index. If a valid device 1046 * with @ifindex is found it will be returned with netdev->lock held. 1047 * netdev_unlock() must be called to release it. 1048 * 1049 * Return: pointer to a device with lock held, NULL if not found. 1050 */ 1051 struct net_device *netdev_get_by_index_lock(struct net *net, int ifindex) 1052 { 1053 struct net_device *dev; 1054 1055 dev = dev_get_by_index(net, ifindex); 1056 if (!dev) 1057 return NULL; 1058 1059 return __netdev_put_lock(dev); 1060 } 1061 1062 /** 1063 * netdev_get_by_name_lock() - find a device by its name 1064 * @net: the applicable net namespace 1065 * @name: name of device 1066 * 1067 * Search for an interface by name. If a valid device 1068 * with @name is found it will be returned with netdev->lock held. 1069 * netdev_unlock() must be called to release it. 1070 * 1071 * Return: pointer to a device with lock held, NULL if not found. 1072 */ 1073 struct net_device *netdev_get_by_name_lock(struct net *net, const char *name) 1074 { 1075 struct net_device *dev; 1076 1077 dev = dev_get_by_name(net, name); 1078 if (!dev) 1079 return NULL; 1080 1081 return __netdev_put_lock(dev); 1082 } 1083 1084 struct net_device * 1085 netdev_xa_find_lock(struct net *net, struct net_device *dev, 1086 unsigned long *index) 1087 { 1088 if (dev) 1089 netdev_unlock(dev); 1090 1091 do { 1092 rcu_read_lock(); 1093 dev = xa_find(&net->dev_by_index, index, ULONG_MAX, XA_PRESENT); 1094 if (!dev) { 1095 rcu_read_unlock(); 1096 return NULL; 1097 } 1098 dev_hold(dev); 1099 rcu_read_unlock(); 1100 1101 dev = __netdev_put_lock(dev); 1102 if (dev) 1103 return dev; 1104 1105 (*index)++; 1106 } while (true); 1107 } 1108 1109 static DEFINE_SEQLOCK(netdev_rename_lock); 1110 1111 void netdev_copy_name(struct net_device *dev, char *name) 1112 { 1113 unsigned int seq; 1114 1115 do { 1116 seq = read_seqbegin(&netdev_rename_lock); 1117 strscpy(name, dev->name, IFNAMSIZ); 1118 } while (read_seqretry(&netdev_rename_lock, seq)); 1119 } 1120 1121 /** 1122 * netdev_get_name - get a netdevice name, knowing its ifindex. 1123 * @net: network namespace 1124 * @name: a pointer to the buffer where the name will be stored. 1125 * @ifindex: the ifindex of the interface to get the name from. 1126 */ 1127 int netdev_get_name(struct net *net, char *name, int ifindex) 1128 { 1129 struct net_device *dev; 1130 int ret; 1131 1132 rcu_read_lock(); 1133 1134 dev = dev_get_by_index_rcu(net, ifindex); 1135 if (!dev) { 1136 ret = -ENODEV; 1137 goto out; 1138 } 1139 1140 netdev_copy_name(dev, name); 1141 1142 ret = 0; 1143 out: 1144 rcu_read_unlock(); 1145 return ret; 1146 } 1147 1148 static bool dev_addr_cmp(struct net_device *dev, unsigned short type, 1149 const char *ha) 1150 { 1151 return dev->type == type && !memcmp(dev->dev_addr, ha, dev->addr_len); 1152 } 1153 1154 /** 1155 * dev_getbyhwaddr_rcu - find a device by its hardware address 1156 * @net: the applicable net namespace 1157 * @type: media type of device 1158 * @ha: hardware address 1159 * 1160 * Search for an interface by MAC address. Returns NULL if the device 1161 * is not found or a pointer to the device. 1162 * The caller must hold RCU. 1163 * The returned device has not had its ref count increased 1164 * and the caller must therefore be careful about locking 1165 * 1166 */ 1167 1168 struct net_device *dev_getbyhwaddr_rcu(struct net *net, unsigned short type, 1169 const char *ha) 1170 { 1171 struct net_device *dev; 1172 1173 for_each_netdev_rcu(net, dev) 1174 if (dev_addr_cmp(dev, type, ha)) 1175 return dev; 1176 1177 return NULL; 1178 } 1179 EXPORT_SYMBOL(dev_getbyhwaddr_rcu); 1180 1181 /** 1182 * dev_getbyhwaddr() - find a device by its hardware address 1183 * @net: the applicable net namespace 1184 * @type: media type of device 1185 * @ha: hardware address 1186 * 1187 * Similar to dev_getbyhwaddr_rcu(), but the owner needs to hold 1188 * rtnl_lock. 1189 * 1190 * Context: rtnl_lock() must be held. 1191 * Return: pointer to the net_device, or NULL if not found 1192 */ 1193 struct net_device *dev_getbyhwaddr(struct net *net, unsigned short type, 1194 const char *ha) 1195 { 1196 struct net_device *dev; 1197 1198 ASSERT_RTNL(); 1199 for_each_netdev(net, dev) 1200 if (dev_addr_cmp(dev, type, ha)) 1201 return dev; 1202 1203 return NULL; 1204 } 1205 EXPORT_SYMBOL(dev_getbyhwaddr); 1206 1207 struct net_device *dev_getfirstbyhwtype(struct net *net, unsigned short type) 1208 { 1209 struct net_device *dev, *ret = NULL; 1210 1211 rcu_read_lock(); 1212 for_each_netdev_rcu(net, dev) 1213 if (dev->type == type) { 1214 dev_hold(dev); 1215 ret = dev; 1216 break; 1217 } 1218 rcu_read_unlock(); 1219 return ret; 1220 } 1221 EXPORT_SYMBOL(dev_getfirstbyhwtype); 1222 1223 /** 1224 * __dev_get_by_flags - find any device with given flags 1225 * @net: the applicable net namespace 1226 * @if_flags: IFF_* values 1227 * @mask: bitmask of bits in if_flags to check 1228 * 1229 * Search for any interface with the given flags. Returns NULL if a device 1230 * is not found or a pointer to the device. Must be called inside 1231 * rtnl_lock(), and result refcount is unchanged. 1232 */ 1233 1234 struct net_device *__dev_get_by_flags(struct net *net, unsigned short if_flags, 1235 unsigned short mask) 1236 { 1237 struct net_device *dev, *ret; 1238 1239 ASSERT_RTNL(); 1240 1241 ret = NULL; 1242 for_each_netdev(net, dev) { 1243 if (((dev->flags ^ if_flags) & mask) == 0) { 1244 ret = dev; 1245 break; 1246 } 1247 } 1248 return ret; 1249 } 1250 EXPORT_SYMBOL(__dev_get_by_flags); 1251 1252 /** 1253 * dev_valid_name - check if name is okay for network device 1254 * @name: name string 1255 * 1256 * Network device names need to be valid file names to 1257 * allow sysfs to work. We also disallow any kind of 1258 * whitespace. 1259 */ 1260 bool dev_valid_name(const char *name) 1261 { 1262 if (*name == '\0') 1263 return false; 1264 if (strnlen(name, IFNAMSIZ) == IFNAMSIZ) 1265 return false; 1266 if (!strcmp(name, ".") || !strcmp(name, "..")) 1267 return false; 1268 1269 while (*name) { 1270 if (*name == '/' || *name == ':' || isspace(*name)) 1271 return false; 1272 name++; 1273 } 1274 return true; 1275 } 1276 EXPORT_SYMBOL(dev_valid_name); 1277 1278 /** 1279 * __dev_alloc_name - allocate a name for a device 1280 * @net: network namespace to allocate the device name in 1281 * @name: name format string 1282 * @res: result name string 1283 * 1284 * Passed a format string - eg "lt%d" it will try and find a suitable 1285 * id. It scans list of devices to build up a free map, then chooses 1286 * the first empty slot. The caller must hold the dev_base or rtnl lock 1287 * while allocating the name and adding the device in order to avoid 1288 * duplicates. 1289 * Limited to bits_per_byte * page size devices (ie 32K on most platforms). 1290 * Returns the number of the unit assigned or a negative errno code. 1291 */ 1292 1293 static int __dev_alloc_name(struct net *net, const char *name, char *res) 1294 { 1295 int i = 0; 1296 const char *p; 1297 const int max_netdevices = 8*PAGE_SIZE; 1298 unsigned long *inuse; 1299 struct net_device *d; 1300 char buf[IFNAMSIZ]; 1301 1302 /* Verify the string as this thing may have come from the user. 1303 * There must be one "%d" and no other "%" characters. 1304 */ 1305 p = strchr(name, '%'); 1306 if (!p || p[1] != 'd' || strchr(p + 2, '%')) 1307 return -EINVAL; 1308 1309 /* Use one page as a bit array of possible slots */ 1310 inuse = bitmap_zalloc(max_netdevices, GFP_ATOMIC); 1311 if (!inuse) 1312 return -ENOMEM; 1313 1314 for_each_netdev(net, d) { 1315 struct netdev_name_node *name_node; 1316 1317 netdev_for_each_altname(d, name_node) { 1318 if (!sscanf(name_node->name, name, &i)) 1319 continue; 1320 if (i < 0 || i >= max_netdevices) 1321 continue; 1322 1323 /* avoid cases where sscanf is not exact inverse of printf */ 1324 snprintf(buf, IFNAMSIZ, name, i); 1325 if (!strncmp(buf, name_node->name, IFNAMSIZ)) 1326 __set_bit(i, inuse); 1327 } 1328 if (!sscanf(d->name, name, &i)) 1329 continue; 1330 if (i < 0 || i >= max_netdevices) 1331 continue; 1332 1333 /* avoid cases where sscanf is not exact inverse of printf */ 1334 snprintf(buf, IFNAMSIZ, name, i); 1335 if (!strncmp(buf, d->name, IFNAMSIZ)) 1336 __set_bit(i, inuse); 1337 } 1338 1339 i = find_first_zero_bit(inuse, max_netdevices); 1340 bitmap_free(inuse); 1341 if (i == max_netdevices) 1342 return -ENFILE; 1343 1344 /* 'res' and 'name' could overlap, use 'buf' as an intermediate buffer */ 1345 strscpy(buf, name, IFNAMSIZ); 1346 snprintf(res, IFNAMSIZ, buf, i); 1347 return i; 1348 } 1349 1350 /* Returns negative errno or allocated unit id (see __dev_alloc_name()) */ 1351 static int dev_prep_valid_name(struct net *net, struct net_device *dev, 1352 const char *want_name, char *out_name, 1353 int dup_errno) 1354 { 1355 if (!dev_valid_name(want_name)) 1356 return -EINVAL; 1357 1358 if (strchr(want_name, '%')) 1359 return __dev_alloc_name(net, want_name, out_name); 1360 1361 if (netdev_name_in_use(net, want_name)) 1362 return -dup_errno; 1363 if (out_name != want_name) 1364 strscpy(out_name, want_name, IFNAMSIZ); 1365 return 0; 1366 } 1367 1368 /** 1369 * dev_alloc_name - allocate a name for a device 1370 * @dev: device 1371 * @name: name format string 1372 * 1373 * Passed a format string - eg "lt%d" it will try and find a suitable 1374 * id. It scans list of devices to build up a free map, then chooses 1375 * the first empty slot. The caller must hold the dev_base or rtnl lock 1376 * while allocating the name and adding the device in order to avoid 1377 * duplicates. 1378 * Limited to bits_per_byte * page size devices (ie 32K on most platforms). 1379 * Returns the number of the unit assigned or a negative errno code. 1380 */ 1381 1382 int dev_alloc_name(struct net_device *dev, const char *name) 1383 { 1384 return dev_prep_valid_name(dev_net(dev), dev, name, dev->name, ENFILE); 1385 } 1386 EXPORT_SYMBOL(dev_alloc_name); 1387 1388 static int dev_get_valid_name(struct net *net, struct net_device *dev, 1389 const char *name) 1390 { 1391 int ret; 1392 1393 ret = dev_prep_valid_name(net, dev, name, dev->name, EEXIST); 1394 return ret < 0 ? ret : 0; 1395 } 1396 1397 int netif_change_name(struct net_device *dev, const char *newname) 1398 { 1399 struct net *net = dev_net(dev); 1400 unsigned char old_assign_type; 1401 char oldname[IFNAMSIZ]; 1402 int err = 0; 1403 int ret; 1404 1405 ASSERT_RTNL_NET(net); 1406 1407 if (!strncmp(newname, dev->name, IFNAMSIZ)) 1408 return 0; 1409 1410 memcpy(oldname, dev->name, IFNAMSIZ); 1411 1412 write_seqlock_bh(&netdev_rename_lock); 1413 err = dev_get_valid_name(net, dev, newname); 1414 write_sequnlock_bh(&netdev_rename_lock); 1415 1416 if (err < 0) 1417 return err; 1418 1419 if (oldname[0] && !strchr(oldname, '%')) 1420 netdev_info(dev, "renamed from %s%s\n", oldname, 1421 dev->flags & IFF_UP ? " (while UP)" : ""); 1422 1423 old_assign_type = dev->name_assign_type; 1424 WRITE_ONCE(dev->name_assign_type, NET_NAME_RENAMED); 1425 1426 rollback: 1427 ret = device_rename(&dev->dev, dev->name); 1428 if (ret) { 1429 write_seqlock_bh(&netdev_rename_lock); 1430 memcpy(dev->name, oldname, IFNAMSIZ); 1431 write_sequnlock_bh(&netdev_rename_lock); 1432 WRITE_ONCE(dev->name_assign_type, old_assign_type); 1433 return ret; 1434 } 1435 1436 netdev_adjacent_rename_links(dev, oldname); 1437 1438 netdev_name_node_del(dev->name_node); 1439 1440 synchronize_net(); 1441 1442 netdev_name_node_add(net, dev->name_node); 1443 1444 ret = call_netdevice_notifiers(NETDEV_CHANGENAME, dev); 1445 ret = notifier_to_errno(ret); 1446 1447 if (ret) { 1448 /* err >= 0 after dev_alloc_name() or stores the first errno */ 1449 if (err >= 0) { 1450 err = ret; 1451 write_seqlock_bh(&netdev_rename_lock); 1452 memcpy(dev->name, oldname, IFNAMSIZ); 1453 write_sequnlock_bh(&netdev_rename_lock); 1454 memcpy(oldname, newname, IFNAMSIZ); 1455 WRITE_ONCE(dev->name_assign_type, old_assign_type); 1456 old_assign_type = NET_NAME_RENAMED; 1457 goto rollback; 1458 } else { 1459 netdev_err(dev, "name change rollback failed: %d\n", 1460 ret); 1461 } 1462 } 1463 1464 return err; 1465 } 1466 1467 int netif_set_alias(struct net_device *dev, const char *alias, size_t len) 1468 { 1469 struct dev_ifalias *new_alias = NULL; 1470 1471 if (len >= IFALIASZ) 1472 return -EINVAL; 1473 1474 if (len) { 1475 new_alias = kmalloc(sizeof(*new_alias) + len + 1, GFP_KERNEL); 1476 if (!new_alias) 1477 return -ENOMEM; 1478 1479 memcpy(new_alias->ifalias, alias, len); 1480 new_alias->ifalias[len] = 0; 1481 } 1482 1483 mutex_lock(&ifalias_mutex); 1484 new_alias = rcu_replace_pointer(dev->ifalias, new_alias, 1485 mutex_is_locked(&ifalias_mutex)); 1486 mutex_unlock(&ifalias_mutex); 1487 1488 if (new_alias) 1489 kfree_rcu(new_alias, rcuhead); 1490 1491 return len; 1492 } 1493 1494 /** 1495 * dev_get_alias - get ifalias of a device 1496 * @dev: device 1497 * @name: buffer to store name of ifalias 1498 * @len: size of buffer 1499 * 1500 * get ifalias for a device. Caller must make sure dev cannot go 1501 * away, e.g. rcu read lock or own a reference count to device. 1502 */ 1503 int dev_get_alias(const struct net_device *dev, char *name, size_t len) 1504 { 1505 const struct dev_ifalias *alias; 1506 int ret = 0; 1507 1508 rcu_read_lock(); 1509 alias = rcu_dereference(dev->ifalias); 1510 if (alias) 1511 ret = snprintf(name, len, "%s", alias->ifalias); 1512 rcu_read_unlock(); 1513 1514 return ret; 1515 } 1516 1517 /** 1518 * netdev_features_change - device changes features 1519 * @dev: device to cause notification 1520 * 1521 * Called to indicate a device has changed features. 1522 */ 1523 void netdev_features_change(struct net_device *dev) 1524 { 1525 call_netdevice_notifiers(NETDEV_FEAT_CHANGE, dev); 1526 } 1527 EXPORT_SYMBOL(netdev_features_change); 1528 1529 /** 1530 * netdev_state_change - device changes state 1531 * @dev: device to cause notification 1532 * 1533 * Called to indicate a device has changed state. This function calls 1534 * the notifier chains for netdev_chain and sends a NEWLINK message 1535 * to the routing socket. 1536 */ 1537 void netdev_state_change(struct net_device *dev) 1538 { 1539 if (dev->flags & IFF_UP) { 1540 struct netdev_notifier_change_info change_info = { 1541 .info.dev = dev, 1542 }; 1543 1544 call_netdevice_notifiers_info(NETDEV_CHANGE, 1545 &change_info.info); 1546 rtmsg_ifinfo(RTM_NEWLINK, dev, 0, GFP_KERNEL, 0, NULL); 1547 } 1548 } 1549 EXPORT_SYMBOL(netdev_state_change); 1550 1551 /** 1552 * __netdev_notify_peers - notify network peers about existence of @dev, 1553 * to be called when rtnl lock is already held. 1554 * @dev: network device 1555 * 1556 * Generate traffic such that interested network peers are aware of 1557 * @dev, such as by generating a gratuitous ARP. This may be used when 1558 * a device wants to inform the rest of the network about some sort of 1559 * reconfiguration such as a failover event or virtual machine 1560 * migration. 1561 */ 1562 void __netdev_notify_peers(struct net_device *dev) 1563 { 1564 ASSERT_RTNL(); 1565 call_netdevice_notifiers(NETDEV_NOTIFY_PEERS, dev); 1566 call_netdevice_notifiers(NETDEV_RESEND_IGMP, dev); 1567 } 1568 EXPORT_SYMBOL(__netdev_notify_peers); 1569 1570 /** 1571 * netdev_notify_peers - notify network peers about existence of @dev 1572 * @dev: network device 1573 * 1574 * Generate traffic such that interested network peers are aware of 1575 * @dev, such as by generating a gratuitous ARP. This may be used when 1576 * a device wants to inform the rest of the network about some sort of 1577 * reconfiguration such as a failover event or virtual machine 1578 * migration. 1579 */ 1580 void netdev_notify_peers(struct net_device *dev) 1581 { 1582 rtnl_lock(); 1583 __netdev_notify_peers(dev); 1584 rtnl_unlock(); 1585 } 1586 EXPORT_SYMBOL(netdev_notify_peers); 1587 1588 static int napi_threaded_poll(void *data); 1589 1590 static int napi_kthread_create(struct napi_struct *n) 1591 { 1592 int err = 0; 1593 1594 /* Create and wake up the kthread once to put it in 1595 * TASK_INTERRUPTIBLE mode to avoid the blocked task 1596 * warning and work with loadavg. 1597 */ 1598 n->thread = kthread_run(napi_threaded_poll, n, "napi/%s-%d", 1599 n->dev->name, n->napi_id); 1600 if (IS_ERR(n->thread)) { 1601 err = PTR_ERR(n->thread); 1602 pr_err("kthread_run failed with err %d\n", err); 1603 n->thread = NULL; 1604 } 1605 1606 return err; 1607 } 1608 1609 static int __dev_open(struct net_device *dev, struct netlink_ext_ack *extack) 1610 { 1611 const struct net_device_ops *ops = dev->netdev_ops; 1612 int ret; 1613 1614 ASSERT_RTNL(); 1615 dev_addr_check(dev); 1616 1617 if (!netif_device_present(dev)) { 1618 /* may be detached because parent is runtime-suspended */ 1619 if (dev->dev.parent) 1620 pm_runtime_resume(dev->dev.parent); 1621 if (!netif_device_present(dev)) 1622 return -ENODEV; 1623 } 1624 1625 /* Block netpoll from trying to do any rx path servicing. 1626 * If we don't do this there is a chance ndo_poll_controller 1627 * or ndo_poll may be running while we open the device 1628 */ 1629 netpoll_poll_disable(dev); 1630 1631 ret = call_netdevice_notifiers_extack(NETDEV_PRE_UP, dev, extack); 1632 ret = notifier_to_errno(ret); 1633 if (ret) 1634 return ret; 1635 1636 set_bit(__LINK_STATE_START, &dev->state); 1637 1638 netdev_ops_assert_locked(dev); 1639 1640 if (ops->ndo_validate_addr) 1641 ret = ops->ndo_validate_addr(dev); 1642 1643 if (!ret && ops->ndo_open) 1644 ret = ops->ndo_open(dev); 1645 1646 netpoll_poll_enable(dev); 1647 1648 if (ret) 1649 clear_bit(__LINK_STATE_START, &dev->state); 1650 else { 1651 netif_set_up(dev, true); 1652 dev_set_rx_mode(dev); 1653 dev_activate(dev); 1654 add_device_randomness(dev->dev_addr, dev->addr_len); 1655 } 1656 1657 return ret; 1658 } 1659 1660 int netif_open(struct net_device *dev, struct netlink_ext_ack *extack) 1661 { 1662 int ret; 1663 1664 if (dev->flags & IFF_UP) 1665 return 0; 1666 1667 ret = __dev_open(dev, extack); 1668 if (ret < 0) 1669 return ret; 1670 1671 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP | IFF_RUNNING, GFP_KERNEL, 0, NULL); 1672 call_netdevice_notifiers(NETDEV_UP, dev); 1673 1674 return ret; 1675 } 1676 1677 static void __dev_close_many(struct list_head *head) 1678 { 1679 struct net_device *dev; 1680 1681 ASSERT_RTNL(); 1682 might_sleep(); 1683 1684 list_for_each_entry(dev, head, close_list) { 1685 /* Temporarily disable netpoll until the interface is down */ 1686 netpoll_poll_disable(dev); 1687 1688 call_netdevice_notifiers(NETDEV_GOING_DOWN, dev); 1689 1690 clear_bit(__LINK_STATE_START, &dev->state); 1691 1692 /* Synchronize to scheduled poll. We cannot touch poll list, it 1693 * can be even on different cpu. So just clear netif_running(). 1694 * 1695 * dev->stop() will invoke napi_disable() on all of it's 1696 * napi_struct instances on this device. 1697 */ 1698 smp_mb__after_atomic(); /* Commit netif_running(). */ 1699 } 1700 1701 dev_deactivate_many(head); 1702 1703 list_for_each_entry(dev, head, close_list) { 1704 const struct net_device_ops *ops = dev->netdev_ops; 1705 1706 /* 1707 * Call the device specific close. This cannot fail. 1708 * Only if device is UP 1709 * 1710 * We allow it to be called even after a DETACH hot-plug 1711 * event. 1712 */ 1713 1714 netdev_ops_assert_locked(dev); 1715 1716 if (ops->ndo_stop) 1717 ops->ndo_stop(dev); 1718 1719 netif_set_up(dev, false); 1720 netpoll_poll_enable(dev); 1721 } 1722 } 1723 1724 static void __dev_close(struct net_device *dev) 1725 { 1726 LIST_HEAD(single); 1727 1728 list_add(&dev->close_list, &single); 1729 __dev_close_many(&single); 1730 list_del(&single); 1731 } 1732 1733 void dev_close_many(struct list_head *head, bool unlink) 1734 { 1735 struct net_device *dev, *tmp; 1736 1737 /* Remove the devices that don't need to be closed */ 1738 list_for_each_entry_safe(dev, tmp, head, close_list) 1739 if (!(dev->flags & IFF_UP)) 1740 list_del_init(&dev->close_list); 1741 1742 __dev_close_many(head); 1743 1744 list_for_each_entry_safe(dev, tmp, head, close_list) { 1745 rtmsg_ifinfo(RTM_NEWLINK, dev, IFF_UP | IFF_RUNNING, GFP_KERNEL, 0, NULL); 1746 call_netdevice_notifiers(NETDEV_DOWN, dev); 1747 if (unlink) 1748 list_del_init(&dev->close_list); 1749 } 1750 } 1751 EXPORT_SYMBOL(dev_close_many); 1752 1753 void netif_close(struct net_device *dev) 1754 { 1755 if (dev->flags & IFF_UP) { 1756 LIST_HEAD(single); 1757 1758 list_add(&dev->close_list, &single); 1759 dev_close_many(&single, true); 1760 list_del(&single); 1761 } 1762 } 1763 EXPORT_SYMBOL(netif_close); 1764 1765 void netif_disable_lro(struct net_device *dev) 1766 { 1767 struct net_device *lower_dev; 1768 struct list_head *iter; 1769 1770 dev->wanted_features &= ~NETIF_F_LRO; 1771 netdev_update_features(dev); 1772 1773 if (unlikely(dev->features & NETIF_F_LRO)) 1774 netdev_WARN(dev, "failed to disable LRO!\n"); 1775 1776 netdev_for_each_lower_dev(dev, lower_dev, iter) { 1777 netdev_lock_ops(lower_dev); 1778 netif_disable_lro(lower_dev); 1779 netdev_unlock_ops(lower_dev); 1780 } 1781 } 1782 1783 /** 1784 * dev_disable_gro_hw - disable HW Generic Receive Offload on a device 1785 * @dev: device 1786 * 1787 * Disable HW Generic Receive Offload (GRO_HW) on a net device. Must be 1788 * called under RTNL. This is needed if Generic XDP is installed on 1789 * the device. 1790 */ 1791 static void dev_disable_gro_hw(struct net_device *dev) 1792 { 1793 dev->wanted_features &= ~NETIF_F_GRO_HW; 1794 netdev_update_features(dev); 1795 1796 if (unlikely(dev->features & NETIF_F_GRO_HW)) 1797 netdev_WARN(dev, "failed to disable GRO_HW!\n"); 1798 } 1799 1800 const char *netdev_cmd_to_name(enum netdev_cmd cmd) 1801 { 1802 #define N(val) \ 1803 case NETDEV_##val: \ 1804 return "NETDEV_" __stringify(val); 1805 switch (cmd) { 1806 N(UP) N(DOWN) N(REBOOT) N(CHANGE) N(REGISTER) N(UNREGISTER) 1807 N(CHANGEMTU) N(CHANGEADDR) N(GOING_DOWN) N(CHANGENAME) N(FEAT_CHANGE) 1808 N(BONDING_FAILOVER) N(PRE_UP) N(PRE_TYPE_CHANGE) N(POST_TYPE_CHANGE) 1809 N(POST_INIT) N(PRE_UNINIT) N(RELEASE) N(NOTIFY_PEERS) N(JOIN) 1810 N(CHANGEUPPER) N(RESEND_IGMP) N(PRECHANGEMTU) N(CHANGEINFODATA) 1811 N(BONDING_INFO) N(PRECHANGEUPPER) N(CHANGELOWERSTATE) 1812 N(UDP_TUNNEL_PUSH_INFO) N(UDP_TUNNEL_DROP_INFO) N(CHANGE_TX_QUEUE_LEN) 1813 N(CVLAN_FILTER_PUSH_INFO) N(CVLAN_FILTER_DROP_INFO) 1814 N(SVLAN_FILTER_PUSH_INFO) N(SVLAN_FILTER_DROP_INFO) 1815 N(PRE_CHANGEADDR) N(OFFLOAD_XSTATS_ENABLE) N(OFFLOAD_XSTATS_DISABLE) 1816 N(OFFLOAD_XSTATS_REPORT_USED) N(OFFLOAD_XSTATS_REPORT_DELTA) 1817 N(XDP_FEAT_CHANGE) 1818 } 1819 #undef N 1820 return "UNKNOWN_NETDEV_EVENT"; 1821 } 1822 EXPORT_SYMBOL_GPL(netdev_cmd_to_name); 1823 1824 static int call_netdevice_notifier(struct notifier_block *nb, unsigned long val, 1825 struct net_device *dev) 1826 { 1827 struct netdev_notifier_info info = { 1828 .dev = dev, 1829 }; 1830 1831 return nb->notifier_call(nb, val, &info); 1832 } 1833 1834 static int call_netdevice_register_notifiers(struct notifier_block *nb, 1835 struct net_device *dev) 1836 { 1837 int err; 1838 1839 err = call_netdevice_notifier(nb, NETDEV_REGISTER, dev); 1840 err = notifier_to_errno(err); 1841 if (err) 1842 return err; 1843 1844 if (!(dev->flags & IFF_UP)) 1845 return 0; 1846 1847 call_netdevice_notifier(nb, NETDEV_UP, dev); 1848 return 0; 1849 } 1850 1851 static void call_netdevice_unregister_notifiers(struct notifier_block *nb, 1852 struct net_device *dev) 1853 { 1854 if (dev->flags & IFF_UP) { 1855 call_netdevice_notifier(nb, NETDEV_GOING_DOWN, 1856 dev); 1857 call_netdevice_notifier(nb, NETDEV_DOWN, dev); 1858 } 1859 call_netdevice_notifier(nb, NETDEV_UNREGISTER, dev); 1860 } 1861 1862 static int call_netdevice_register_net_notifiers(struct notifier_block *nb, 1863 struct net *net) 1864 { 1865 struct net_device *dev; 1866 int err; 1867 1868 for_each_netdev(net, dev) { 1869 err = call_netdevice_register_notifiers(nb, dev); 1870 if (err) 1871 goto rollback; 1872 } 1873 return 0; 1874 1875 rollback: 1876 for_each_netdev_continue_reverse(net, dev) 1877 call_netdevice_unregister_notifiers(nb, dev); 1878 return err; 1879 } 1880 1881 static void call_netdevice_unregister_net_notifiers(struct notifier_block *nb, 1882 struct net *net) 1883 { 1884 struct net_device *dev; 1885 1886 for_each_netdev(net, dev) 1887 call_netdevice_unregister_notifiers(nb, dev); 1888 } 1889 1890 static int dev_boot_phase = 1; 1891 1892 /** 1893 * register_netdevice_notifier - register a network notifier block 1894 * @nb: notifier 1895 * 1896 * Register a notifier to be called when network device events occur. 1897 * The notifier passed is linked into the kernel structures and must 1898 * not be reused until it has been unregistered. A negative errno code 1899 * is returned on a failure. 1900 * 1901 * When registered all registration and up events are replayed 1902 * to the new notifier to allow device to have a race free 1903 * view of the network device list. 1904 */ 1905 1906 int register_netdevice_notifier(struct notifier_block *nb) 1907 { 1908 struct net *net; 1909 int err; 1910 1911 /* Close race with setup_net() and cleanup_net() */ 1912 down_write(&pernet_ops_rwsem); 1913 1914 /* When RTNL is removed, we need protection for netdev_chain. */ 1915 rtnl_lock(); 1916 1917 err = raw_notifier_chain_register(&netdev_chain, nb); 1918 if (err) 1919 goto unlock; 1920 if (dev_boot_phase) 1921 goto unlock; 1922 for_each_net(net) { 1923 __rtnl_net_lock(net); 1924 err = call_netdevice_register_net_notifiers(nb, net); 1925 __rtnl_net_unlock(net); 1926 if (err) 1927 goto rollback; 1928 } 1929 1930 unlock: 1931 rtnl_unlock(); 1932 up_write(&pernet_ops_rwsem); 1933 return err; 1934 1935 rollback: 1936 for_each_net_continue_reverse(net) { 1937 __rtnl_net_lock(net); 1938 call_netdevice_unregister_net_notifiers(nb, net); 1939 __rtnl_net_unlock(net); 1940 } 1941 1942 raw_notifier_chain_unregister(&netdev_chain, nb); 1943 goto unlock; 1944 } 1945 EXPORT_SYMBOL(register_netdevice_notifier); 1946 1947 /** 1948 * unregister_netdevice_notifier - unregister a network notifier block 1949 * @nb: notifier 1950 * 1951 * Unregister a notifier previously registered by 1952 * register_netdevice_notifier(). The notifier is unlinked into the 1953 * kernel structures and may then be reused. A negative errno code 1954 * is returned on a failure. 1955 * 1956 * After unregistering unregister and down device events are synthesized 1957 * for all devices on the device list to the removed notifier to remove 1958 * the need for special case cleanup code. 1959 */ 1960 1961 int unregister_netdevice_notifier(struct notifier_block *nb) 1962 { 1963 struct net *net; 1964 int err; 1965 1966 /* Close race with setup_net() and cleanup_net() */ 1967 down_write(&pernet_ops_rwsem); 1968 rtnl_lock(); 1969 err = raw_notifier_chain_unregister(&netdev_chain, nb); 1970 if (err) 1971 goto unlock; 1972 1973 for_each_net(net) { 1974 __rtnl_net_lock(net); 1975 call_netdevice_unregister_net_notifiers(nb, net); 1976 __rtnl_net_unlock(net); 1977 } 1978 1979 unlock: 1980 rtnl_unlock(); 1981 up_write(&pernet_ops_rwsem); 1982 return err; 1983 } 1984 EXPORT_SYMBOL(unregister_netdevice_notifier); 1985 1986 static int __register_netdevice_notifier_net(struct net *net, 1987 struct notifier_block *nb, 1988 bool ignore_call_fail) 1989 { 1990 int err; 1991 1992 err = raw_notifier_chain_register(&net->netdev_chain, nb); 1993 if (err) 1994 return err; 1995 if (dev_boot_phase) 1996 return 0; 1997 1998 err = call_netdevice_register_net_notifiers(nb, net); 1999 if (err && !ignore_call_fail) 2000 goto chain_unregister; 2001 2002 return 0; 2003 2004 chain_unregister: 2005 raw_notifier_chain_unregister(&net->netdev_chain, nb); 2006 return err; 2007 } 2008 2009 static int __unregister_netdevice_notifier_net(struct net *net, 2010 struct notifier_block *nb) 2011 { 2012 int err; 2013 2014 err = raw_notifier_chain_unregister(&net->netdev_chain, nb); 2015 if (err) 2016 return err; 2017 2018 call_netdevice_unregister_net_notifiers(nb, net); 2019 return 0; 2020 } 2021 2022 /** 2023 * register_netdevice_notifier_net - register a per-netns network notifier block 2024 * @net: network namespace 2025 * @nb: notifier 2026 * 2027 * Register a notifier to be called when network device events occur. 2028 * The notifier passed is linked into the kernel structures and must 2029 * not be reused until it has been unregistered. A negative errno code 2030 * is returned on a failure. 2031 * 2032 * When registered all registration and up events are replayed 2033 * to the new notifier to allow device to have a race free 2034 * view of the network device list. 2035 */ 2036 2037 int register_netdevice_notifier_net(struct net *net, struct notifier_block *nb) 2038 { 2039 int err; 2040 2041 rtnl_net_lock(net); 2042 err = __register_netdevice_notifier_net(net, nb, false); 2043 rtnl_net_unlock(net); 2044 2045 return err; 2046 } 2047 EXPORT_SYMBOL(register_netdevice_notifier_net); 2048 2049 /** 2050 * unregister_netdevice_notifier_net - unregister a per-netns 2051 * network notifier block 2052 * @net: network namespace 2053 * @nb: notifier 2054 * 2055 * Unregister a notifier previously registered by 2056 * register_netdevice_notifier_net(). The notifier is unlinked from the 2057 * kernel structures and may then be reused. A negative errno code 2058 * is returned on a failure. 2059 * 2060 * After unregistering unregister and down device events are synthesized 2061 * for all devices on the device list to the removed notifier to remove 2062 * the need for special case cleanup code. 2063 */ 2064 2065 int unregister_netdevice_notifier_net(struct net *net, 2066 struct notifier_block *nb) 2067 { 2068 int err; 2069 2070 rtnl_net_lock(net); 2071 err = __unregister_netdevice_notifier_net(net, nb); 2072 rtnl_net_unlock(net); 2073 2074 return err; 2075 } 2076 EXPORT_SYMBOL(unregister_netdevice_notifier_net); 2077 2078 static void __move_netdevice_notifier_net(struct net *src_net, 2079 struct net *dst_net, 2080 struct notifier_block *nb) 2081 { 2082 __unregister_netdevice_notifier_net(src_net, nb); 2083 __register_netdevice_notifier_net(dst_net, nb, true); 2084 } 2085 2086 static void rtnl_net_dev_lock(struct net_device *dev) 2087 { 2088 bool again; 2089 2090 do { 2091 struct net *net; 2092 2093 again = false; 2094 2095 /* netns might be being dismantled. */ 2096 rcu_read_lock(); 2097 net = dev_net_rcu(dev); 2098 net_passive_inc(net); 2099 rcu_read_unlock(); 2100 2101 rtnl_net_lock(net); 2102 2103 #ifdef CONFIG_NET_NS 2104 /* dev might have been moved to another netns. */ 2105 if (!net_eq(net, rcu_access_pointer(dev->nd_net.net))) { 2106 rtnl_net_unlock(net); 2107 net_passive_dec(net); 2108 again = true; 2109 } 2110 #endif 2111 } while (again); 2112 } 2113 2114 static void rtnl_net_dev_unlock(struct net_device *dev) 2115 { 2116 struct net *net = dev_net(dev); 2117 2118 rtnl_net_unlock(net); 2119 net_passive_dec(net); 2120 } 2121 2122 int register_netdevice_notifier_dev_net(struct net_device *dev, 2123 struct notifier_block *nb, 2124 struct netdev_net_notifier *nn) 2125 { 2126 int err; 2127 2128 rtnl_net_dev_lock(dev); 2129 err = __register_netdevice_notifier_net(dev_net(dev), nb, false); 2130 if (!err) { 2131 nn->nb = nb; 2132 list_add(&nn->list, &dev->net_notifier_list); 2133 } 2134 rtnl_net_dev_unlock(dev); 2135 2136 return err; 2137 } 2138 EXPORT_SYMBOL(register_netdevice_notifier_dev_net); 2139 2140 int unregister_netdevice_notifier_dev_net(struct net_device *dev, 2141 struct notifier_block *nb, 2142 struct netdev_net_notifier *nn) 2143 { 2144 int err; 2145 2146 rtnl_net_dev_lock(dev); 2147 list_del(&nn->list); 2148 err = __unregister_netdevice_notifier_net(dev_net(dev), nb); 2149 rtnl_net_dev_unlock(dev); 2150 2151 return err; 2152 } 2153 EXPORT_SYMBOL(unregister_netdevice_notifier_dev_net); 2154 2155 static void move_netdevice_notifiers_dev_net(struct net_device *dev, 2156 struct net *net) 2157 { 2158 struct netdev_net_notifier *nn; 2159 2160 list_for_each_entry(nn, &dev->net_notifier_list, list) 2161 __move_netdevice_notifier_net(dev_net(dev), net, nn->nb); 2162 } 2163 2164 /** 2165 * call_netdevice_notifiers_info - call all network notifier blocks 2166 * @val: value passed unmodified to notifier function 2167 * @info: notifier information data 2168 * 2169 * Call all network notifier blocks. Parameters and return value 2170 * are as for raw_notifier_call_chain(). 2171 */ 2172 2173 int call_netdevice_notifiers_info(unsigned long val, 2174 struct netdev_notifier_info *info) 2175 { 2176 struct net *net = dev_net(info->dev); 2177 int ret; 2178 2179 ASSERT_RTNL(); 2180 2181 /* Run per-netns notifier block chain first, then run the global one. 2182 * Hopefully, one day, the global one is going to be removed after 2183 * all notifier block registrators get converted to be per-netns. 2184 */ 2185 ret = raw_notifier_call_chain(&net->netdev_chain, val, info); 2186 if (ret & NOTIFY_STOP_MASK) 2187 return ret; 2188 return raw_notifier_call_chain(&netdev_chain, val, info); 2189 } 2190 2191 /** 2192 * call_netdevice_notifiers_info_robust - call per-netns notifier blocks 2193 * for and rollback on error 2194 * @val_up: value passed unmodified to notifier function 2195 * @val_down: value passed unmodified to the notifier function when 2196 * recovering from an error on @val_up 2197 * @info: notifier information data 2198 * 2199 * Call all per-netns network notifier blocks, but not notifier blocks on 2200 * the global notifier chain. Parameters and return value are as for 2201 * raw_notifier_call_chain_robust(). 2202 */ 2203 2204 static int 2205 call_netdevice_notifiers_info_robust(unsigned long val_up, 2206 unsigned long val_down, 2207 struct netdev_notifier_info *info) 2208 { 2209 struct net *net = dev_net(info->dev); 2210 2211 ASSERT_RTNL(); 2212 2213 return raw_notifier_call_chain_robust(&net->netdev_chain, 2214 val_up, val_down, info); 2215 } 2216 2217 static int call_netdevice_notifiers_extack(unsigned long val, 2218 struct net_device *dev, 2219 struct netlink_ext_ack *extack) 2220 { 2221 struct netdev_notifier_info info = { 2222 .dev = dev, 2223 .extack = extack, 2224 }; 2225 2226 return call_netdevice_notifiers_info(val, &info); 2227 } 2228 2229 /** 2230 * call_netdevice_notifiers - call all network notifier blocks 2231 * @val: value passed unmodified to notifier function 2232 * @dev: net_device pointer passed unmodified to notifier function 2233 * 2234 * Call all network notifier blocks. Parameters and return value 2235 * are as for raw_notifier_call_chain(). 2236 */ 2237 2238 int call_netdevice_notifiers(unsigned long val, struct net_device *dev) 2239 { 2240 return call_netdevice_notifiers_extack(val, dev, NULL); 2241 } 2242 EXPORT_SYMBOL(call_netdevice_notifiers); 2243 2244 /** 2245 * call_netdevice_notifiers_mtu - call all network notifier blocks 2246 * @val: value passed unmodified to notifier function 2247 * @dev: net_device pointer passed unmodified to notifier function 2248 * @arg: additional u32 argument passed to the notifier function 2249 * 2250 * Call all network notifier blocks. Parameters and return value 2251 * are as for raw_notifier_call_chain(). 2252 */ 2253 static int call_netdevice_notifiers_mtu(unsigned long val, 2254 struct net_device *dev, u32 arg) 2255 { 2256 struct netdev_notifier_info_ext info = { 2257 .info.dev = dev, 2258 .ext.mtu = arg, 2259 }; 2260 2261 BUILD_BUG_ON(offsetof(struct netdev_notifier_info_ext, info) != 0); 2262 2263 return call_netdevice_notifiers_info(val, &info.info); 2264 } 2265 2266 #ifdef CONFIG_NET_INGRESS 2267 static DEFINE_STATIC_KEY_FALSE(ingress_needed_key); 2268 2269 void net_inc_ingress_queue(void) 2270 { 2271 static_branch_inc(&ingress_needed_key); 2272 } 2273 EXPORT_SYMBOL_GPL(net_inc_ingress_queue); 2274 2275 void net_dec_ingress_queue(void) 2276 { 2277 static_branch_dec(&ingress_needed_key); 2278 } 2279 EXPORT_SYMBOL_GPL(net_dec_ingress_queue); 2280 #endif 2281 2282 #ifdef CONFIG_NET_EGRESS 2283 static DEFINE_STATIC_KEY_FALSE(egress_needed_key); 2284 2285 void net_inc_egress_queue(void) 2286 { 2287 static_branch_inc(&egress_needed_key); 2288 } 2289 EXPORT_SYMBOL_GPL(net_inc_egress_queue); 2290 2291 void net_dec_egress_queue(void) 2292 { 2293 static_branch_dec(&egress_needed_key); 2294 } 2295 EXPORT_SYMBOL_GPL(net_dec_egress_queue); 2296 #endif 2297 2298 #ifdef CONFIG_NET_CLS_ACT 2299 DEFINE_STATIC_KEY_FALSE(tcf_sw_enabled_key); 2300 EXPORT_SYMBOL(tcf_sw_enabled_key); 2301 #endif 2302 2303 DEFINE_STATIC_KEY_FALSE(netstamp_needed_key); 2304 EXPORT_SYMBOL(netstamp_needed_key); 2305 #ifdef CONFIG_JUMP_LABEL 2306 static atomic_t netstamp_needed_deferred; 2307 static atomic_t netstamp_wanted; 2308 static void netstamp_clear(struct work_struct *work) 2309 { 2310 int deferred = atomic_xchg(&netstamp_needed_deferred, 0); 2311 int wanted; 2312 2313 wanted = atomic_add_return(deferred, &netstamp_wanted); 2314 if (wanted > 0) 2315 static_branch_enable(&netstamp_needed_key); 2316 else 2317 static_branch_disable(&netstamp_needed_key); 2318 } 2319 static DECLARE_WORK(netstamp_work, netstamp_clear); 2320 #endif 2321 2322 void net_enable_timestamp(void) 2323 { 2324 #ifdef CONFIG_JUMP_LABEL 2325 int wanted = atomic_read(&netstamp_wanted); 2326 2327 while (wanted > 0) { 2328 if (atomic_try_cmpxchg(&netstamp_wanted, &wanted, wanted + 1)) 2329 return; 2330 } 2331 atomic_inc(&netstamp_needed_deferred); 2332 schedule_work(&netstamp_work); 2333 #else 2334 static_branch_inc(&netstamp_needed_key); 2335 #endif 2336 } 2337 EXPORT_SYMBOL(net_enable_timestamp); 2338 2339 void net_disable_timestamp(void) 2340 { 2341 #ifdef CONFIG_JUMP_LABEL 2342 int wanted = atomic_read(&netstamp_wanted); 2343 2344 while (wanted > 1) { 2345 if (atomic_try_cmpxchg(&netstamp_wanted, &wanted, wanted - 1)) 2346 return; 2347 } 2348 atomic_dec(&netstamp_needed_deferred); 2349 schedule_work(&netstamp_work); 2350 #else 2351 static_branch_dec(&netstamp_needed_key); 2352 #endif 2353 } 2354 EXPORT_SYMBOL(net_disable_timestamp); 2355 2356 static inline void net_timestamp_set(struct sk_buff *skb) 2357 { 2358 skb->tstamp = 0; 2359 skb->tstamp_type = SKB_CLOCK_REALTIME; 2360 if (static_branch_unlikely(&netstamp_needed_key)) 2361 skb->tstamp = ktime_get_real(); 2362 } 2363 2364 #define net_timestamp_check(COND, SKB) \ 2365 if (static_branch_unlikely(&netstamp_needed_key)) { \ 2366 if ((COND) && !(SKB)->tstamp) \ 2367 (SKB)->tstamp = ktime_get_real(); \ 2368 } \ 2369 2370 bool is_skb_forwardable(const struct net_device *dev, const struct sk_buff *skb) 2371 { 2372 return __is_skb_forwardable(dev, skb, true); 2373 } 2374 EXPORT_SYMBOL_GPL(is_skb_forwardable); 2375 2376 static int __dev_forward_skb2(struct net_device *dev, struct sk_buff *skb, 2377 bool check_mtu) 2378 { 2379 int ret = ____dev_forward_skb(dev, skb, check_mtu); 2380 2381 if (likely(!ret)) { 2382 skb->protocol = eth_type_trans(skb, dev); 2383 skb_postpull_rcsum(skb, eth_hdr(skb), ETH_HLEN); 2384 } 2385 2386 return ret; 2387 } 2388 2389 int __dev_forward_skb(struct net_device *dev, struct sk_buff *skb) 2390 { 2391 return __dev_forward_skb2(dev, skb, true); 2392 } 2393 EXPORT_SYMBOL_GPL(__dev_forward_skb); 2394 2395 /** 2396 * dev_forward_skb - loopback an skb to another netif 2397 * 2398 * @dev: destination network device 2399 * @skb: buffer to forward 2400 * 2401 * return values: 2402 * NET_RX_SUCCESS (no congestion) 2403 * NET_RX_DROP (packet was dropped, but freed) 2404 * 2405 * dev_forward_skb can be used for injecting an skb from the 2406 * start_xmit function of one device into the receive queue 2407 * of another device. 2408 * 2409 * The receiving device may be in another namespace, so 2410 * we have to clear all information in the skb that could 2411 * impact namespace isolation. 2412 */ 2413 int dev_forward_skb(struct net_device *dev, struct sk_buff *skb) 2414 { 2415 return __dev_forward_skb(dev, skb) ?: netif_rx_internal(skb); 2416 } 2417 EXPORT_SYMBOL_GPL(dev_forward_skb); 2418 2419 int dev_forward_skb_nomtu(struct net_device *dev, struct sk_buff *skb) 2420 { 2421 return __dev_forward_skb2(dev, skb, false) ?: netif_rx_internal(skb); 2422 } 2423 2424 static inline int deliver_skb(struct sk_buff *skb, 2425 struct packet_type *pt_prev, 2426 struct net_device *orig_dev) 2427 { 2428 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC))) 2429 return -ENOMEM; 2430 refcount_inc(&skb->users); 2431 return pt_prev->func(skb, skb->dev, pt_prev, orig_dev); 2432 } 2433 2434 static inline void deliver_ptype_list_skb(struct sk_buff *skb, 2435 struct packet_type **pt, 2436 struct net_device *orig_dev, 2437 __be16 type, 2438 struct list_head *ptype_list) 2439 { 2440 struct packet_type *ptype, *pt_prev = *pt; 2441 2442 list_for_each_entry_rcu(ptype, ptype_list, list) { 2443 if (ptype->type != type) 2444 continue; 2445 if (pt_prev) 2446 deliver_skb(skb, pt_prev, orig_dev); 2447 pt_prev = ptype; 2448 } 2449 *pt = pt_prev; 2450 } 2451 2452 static inline bool skb_loop_sk(struct packet_type *ptype, struct sk_buff *skb) 2453 { 2454 if (!ptype->af_packet_priv || !skb->sk) 2455 return false; 2456 2457 if (ptype->id_match) 2458 return ptype->id_match(ptype, skb->sk); 2459 else if ((struct sock *)ptype->af_packet_priv == skb->sk) 2460 return true; 2461 2462 return false; 2463 } 2464 2465 /** 2466 * dev_nit_active - return true if any network interface taps are in use 2467 * 2468 * @dev: network device to check for the presence of taps 2469 */ 2470 bool dev_nit_active(struct net_device *dev) 2471 { 2472 return !list_empty(&net_hotdata.ptype_all) || 2473 !list_empty(&dev->ptype_all); 2474 } 2475 EXPORT_SYMBOL_GPL(dev_nit_active); 2476 2477 /* 2478 * Support routine. Sends outgoing frames to any network 2479 * taps currently in use. 2480 */ 2481 2482 void dev_queue_xmit_nit(struct sk_buff *skb, struct net_device *dev) 2483 { 2484 struct list_head *ptype_list = &net_hotdata.ptype_all; 2485 struct packet_type *ptype, *pt_prev = NULL; 2486 struct sk_buff *skb2 = NULL; 2487 2488 rcu_read_lock(); 2489 again: 2490 list_for_each_entry_rcu(ptype, ptype_list, list) { 2491 if (READ_ONCE(ptype->ignore_outgoing)) 2492 continue; 2493 2494 /* Never send packets back to the socket 2495 * they originated from - MvS (miquels@drinkel.ow.org) 2496 */ 2497 if (skb_loop_sk(ptype, skb)) 2498 continue; 2499 2500 if (pt_prev) { 2501 deliver_skb(skb2, pt_prev, skb->dev); 2502 pt_prev = ptype; 2503 continue; 2504 } 2505 2506 /* need to clone skb, done only once */ 2507 skb2 = skb_clone(skb, GFP_ATOMIC); 2508 if (!skb2) 2509 goto out_unlock; 2510 2511 net_timestamp_set(skb2); 2512 2513 /* skb->nh should be correctly 2514 * set by sender, so that the second statement is 2515 * just protection against buggy protocols. 2516 */ 2517 skb_reset_mac_header(skb2); 2518 2519 if (skb_network_header(skb2) < skb2->data || 2520 skb_network_header(skb2) > skb_tail_pointer(skb2)) { 2521 net_crit_ratelimited("protocol %04x is buggy, dev %s\n", 2522 ntohs(skb2->protocol), 2523 dev->name); 2524 skb_reset_network_header(skb2); 2525 } 2526 2527 skb2->transport_header = skb2->network_header; 2528 skb2->pkt_type = PACKET_OUTGOING; 2529 pt_prev = ptype; 2530 } 2531 2532 if (ptype_list == &net_hotdata.ptype_all) { 2533 ptype_list = &dev->ptype_all; 2534 goto again; 2535 } 2536 out_unlock: 2537 if (pt_prev) { 2538 if (!skb_orphan_frags_rx(skb2, GFP_ATOMIC)) 2539 pt_prev->func(skb2, skb->dev, pt_prev, skb->dev); 2540 else 2541 kfree_skb(skb2); 2542 } 2543 rcu_read_unlock(); 2544 } 2545 EXPORT_SYMBOL_GPL(dev_queue_xmit_nit); 2546 2547 /** 2548 * netif_setup_tc - Handle tc mappings on real_num_tx_queues change 2549 * @dev: Network device 2550 * @txq: number of queues available 2551 * 2552 * If real_num_tx_queues is changed the tc mappings may no longer be 2553 * valid. To resolve this verify the tc mapping remains valid and if 2554 * not NULL the mapping. With no priorities mapping to this 2555 * offset/count pair it will no longer be used. In the worst case TC0 2556 * is invalid nothing can be done so disable priority mappings. If is 2557 * expected that drivers will fix this mapping if they can before 2558 * calling netif_set_real_num_tx_queues. 2559 */ 2560 static void netif_setup_tc(struct net_device *dev, unsigned int txq) 2561 { 2562 int i; 2563 struct netdev_tc_txq *tc = &dev->tc_to_txq[0]; 2564 2565 /* If TC0 is invalidated disable TC mapping */ 2566 if (tc->offset + tc->count > txq) { 2567 netdev_warn(dev, "Number of in use tx queues changed invalidating tc mappings. Priority traffic classification disabled!\n"); 2568 dev->num_tc = 0; 2569 return; 2570 } 2571 2572 /* Invalidated prio to tc mappings set to TC0 */ 2573 for (i = 1; i < TC_BITMASK + 1; i++) { 2574 int q = netdev_get_prio_tc_map(dev, i); 2575 2576 tc = &dev->tc_to_txq[q]; 2577 if (tc->offset + tc->count > txq) { 2578 netdev_warn(dev, "Number of in use tx queues changed. Priority %i to tc mapping %i is no longer valid. Setting map to 0\n", 2579 i, q); 2580 netdev_set_prio_tc_map(dev, i, 0); 2581 } 2582 } 2583 } 2584 2585 int netdev_txq_to_tc(struct net_device *dev, unsigned int txq) 2586 { 2587 if (dev->num_tc) { 2588 struct netdev_tc_txq *tc = &dev->tc_to_txq[0]; 2589 int i; 2590 2591 /* walk through the TCs and see if it falls into any of them */ 2592 for (i = 0; i < TC_MAX_QUEUE; i++, tc++) { 2593 if ((txq - tc->offset) < tc->count) 2594 return i; 2595 } 2596 2597 /* didn't find it, just return -1 to indicate no match */ 2598 return -1; 2599 } 2600 2601 return 0; 2602 } 2603 EXPORT_SYMBOL(netdev_txq_to_tc); 2604 2605 #ifdef CONFIG_XPS 2606 static struct static_key xps_needed __read_mostly; 2607 static struct static_key xps_rxqs_needed __read_mostly; 2608 static DEFINE_MUTEX(xps_map_mutex); 2609 #define xmap_dereference(P) \ 2610 rcu_dereference_protected((P), lockdep_is_held(&xps_map_mutex)) 2611 2612 static bool remove_xps_queue(struct xps_dev_maps *dev_maps, 2613 struct xps_dev_maps *old_maps, int tci, u16 index) 2614 { 2615 struct xps_map *map = NULL; 2616 int pos; 2617 2618 map = xmap_dereference(dev_maps->attr_map[tci]); 2619 if (!map) 2620 return false; 2621 2622 for (pos = map->len; pos--;) { 2623 if (map->queues[pos] != index) 2624 continue; 2625 2626 if (map->len > 1) { 2627 map->queues[pos] = map->queues[--map->len]; 2628 break; 2629 } 2630 2631 if (old_maps) 2632 RCU_INIT_POINTER(old_maps->attr_map[tci], NULL); 2633 RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL); 2634 kfree_rcu(map, rcu); 2635 return false; 2636 } 2637 2638 return true; 2639 } 2640 2641 static bool remove_xps_queue_cpu(struct net_device *dev, 2642 struct xps_dev_maps *dev_maps, 2643 int cpu, u16 offset, u16 count) 2644 { 2645 int num_tc = dev_maps->num_tc; 2646 bool active = false; 2647 int tci; 2648 2649 for (tci = cpu * num_tc; num_tc--; tci++) { 2650 int i, j; 2651 2652 for (i = count, j = offset; i--; j++) { 2653 if (!remove_xps_queue(dev_maps, NULL, tci, j)) 2654 break; 2655 } 2656 2657 active |= i < 0; 2658 } 2659 2660 return active; 2661 } 2662 2663 static void reset_xps_maps(struct net_device *dev, 2664 struct xps_dev_maps *dev_maps, 2665 enum xps_map_type type) 2666 { 2667 static_key_slow_dec_cpuslocked(&xps_needed); 2668 if (type == XPS_RXQS) 2669 static_key_slow_dec_cpuslocked(&xps_rxqs_needed); 2670 2671 RCU_INIT_POINTER(dev->xps_maps[type], NULL); 2672 2673 kfree_rcu(dev_maps, rcu); 2674 } 2675 2676 static void clean_xps_maps(struct net_device *dev, enum xps_map_type type, 2677 u16 offset, u16 count) 2678 { 2679 struct xps_dev_maps *dev_maps; 2680 bool active = false; 2681 int i, j; 2682 2683 dev_maps = xmap_dereference(dev->xps_maps[type]); 2684 if (!dev_maps) 2685 return; 2686 2687 for (j = 0; j < dev_maps->nr_ids; j++) 2688 active |= remove_xps_queue_cpu(dev, dev_maps, j, offset, count); 2689 if (!active) 2690 reset_xps_maps(dev, dev_maps, type); 2691 2692 if (type == XPS_CPUS) { 2693 for (i = offset + (count - 1); count--; i--) 2694 netdev_queue_numa_node_write( 2695 netdev_get_tx_queue(dev, i), NUMA_NO_NODE); 2696 } 2697 } 2698 2699 static void netif_reset_xps_queues(struct net_device *dev, u16 offset, 2700 u16 count) 2701 { 2702 if (!static_key_false(&xps_needed)) 2703 return; 2704 2705 cpus_read_lock(); 2706 mutex_lock(&xps_map_mutex); 2707 2708 if (static_key_false(&xps_rxqs_needed)) 2709 clean_xps_maps(dev, XPS_RXQS, offset, count); 2710 2711 clean_xps_maps(dev, XPS_CPUS, offset, count); 2712 2713 mutex_unlock(&xps_map_mutex); 2714 cpus_read_unlock(); 2715 } 2716 2717 static void netif_reset_xps_queues_gt(struct net_device *dev, u16 index) 2718 { 2719 netif_reset_xps_queues(dev, index, dev->num_tx_queues - index); 2720 } 2721 2722 static struct xps_map *expand_xps_map(struct xps_map *map, int attr_index, 2723 u16 index, bool is_rxqs_map) 2724 { 2725 struct xps_map *new_map; 2726 int alloc_len = XPS_MIN_MAP_ALLOC; 2727 int i, pos; 2728 2729 for (pos = 0; map && pos < map->len; pos++) { 2730 if (map->queues[pos] != index) 2731 continue; 2732 return map; 2733 } 2734 2735 /* Need to add tx-queue to this CPU's/rx-queue's existing map */ 2736 if (map) { 2737 if (pos < map->alloc_len) 2738 return map; 2739 2740 alloc_len = map->alloc_len * 2; 2741 } 2742 2743 /* Need to allocate new map to store tx-queue on this CPU's/rx-queue's 2744 * map 2745 */ 2746 if (is_rxqs_map) 2747 new_map = kzalloc(XPS_MAP_SIZE(alloc_len), GFP_KERNEL); 2748 else 2749 new_map = kzalloc_node(XPS_MAP_SIZE(alloc_len), GFP_KERNEL, 2750 cpu_to_node(attr_index)); 2751 if (!new_map) 2752 return NULL; 2753 2754 for (i = 0; i < pos; i++) 2755 new_map->queues[i] = map->queues[i]; 2756 new_map->alloc_len = alloc_len; 2757 new_map->len = pos; 2758 2759 return new_map; 2760 } 2761 2762 /* Copy xps maps at a given index */ 2763 static void xps_copy_dev_maps(struct xps_dev_maps *dev_maps, 2764 struct xps_dev_maps *new_dev_maps, int index, 2765 int tc, bool skip_tc) 2766 { 2767 int i, tci = index * dev_maps->num_tc; 2768 struct xps_map *map; 2769 2770 /* copy maps belonging to foreign traffic classes */ 2771 for (i = 0; i < dev_maps->num_tc; i++, tci++) { 2772 if (i == tc && skip_tc) 2773 continue; 2774 2775 /* fill in the new device map from the old device map */ 2776 map = xmap_dereference(dev_maps->attr_map[tci]); 2777 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map); 2778 } 2779 } 2780 2781 /* Must be called under cpus_read_lock */ 2782 int __netif_set_xps_queue(struct net_device *dev, const unsigned long *mask, 2783 u16 index, enum xps_map_type type) 2784 { 2785 struct xps_dev_maps *dev_maps, *new_dev_maps = NULL, *old_dev_maps = NULL; 2786 const unsigned long *online_mask = NULL; 2787 bool active = false, copy = false; 2788 int i, j, tci, numa_node_id = -2; 2789 int maps_sz, num_tc = 1, tc = 0; 2790 struct xps_map *map, *new_map; 2791 unsigned int nr_ids; 2792 2793 WARN_ON_ONCE(index >= dev->num_tx_queues); 2794 2795 if (dev->num_tc) { 2796 /* Do not allow XPS on subordinate device directly */ 2797 num_tc = dev->num_tc; 2798 if (num_tc < 0) 2799 return -EINVAL; 2800 2801 /* If queue belongs to subordinate dev use its map */ 2802 dev = netdev_get_tx_queue(dev, index)->sb_dev ? : dev; 2803 2804 tc = netdev_txq_to_tc(dev, index); 2805 if (tc < 0) 2806 return -EINVAL; 2807 } 2808 2809 mutex_lock(&xps_map_mutex); 2810 2811 dev_maps = xmap_dereference(dev->xps_maps[type]); 2812 if (type == XPS_RXQS) { 2813 maps_sz = XPS_RXQ_DEV_MAPS_SIZE(num_tc, dev->num_rx_queues); 2814 nr_ids = dev->num_rx_queues; 2815 } else { 2816 maps_sz = XPS_CPU_DEV_MAPS_SIZE(num_tc); 2817 if (num_possible_cpus() > 1) 2818 online_mask = cpumask_bits(cpu_online_mask); 2819 nr_ids = nr_cpu_ids; 2820 } 2821 2822 if (maps_sz < L1_CACHE_BYTES) 2823 maps_sz = L1_CACHE_BYTES; 2824 2825 /* The old dev_maps could be larger or smaller than the one we're 2826 * setting up now, as dev->num_tc or nr_ids could have been updated in 2827 * between. We could try to be smart, but let's be safe instead and only 2828 * copy foreign traffic classes if the two map sizes match. 2829 */ 2830 if (dev_maps && 2831 dev_maps->num_tc == num_tc && dev_maps->nr_ids == nr_ids) 2832 copy = true; 2833 2834 /* allocate memory for queue storage */ 2835 for (j = -1; j = netif_attrmask_next_and(j, online_mask, mask, nr_ids), 2836 j < nr_ids;) { 2837 if (!new_dev_maps) { 2838 new_dev_maps = kzalloc(maps_sz, GFP_KERNEL); 2839 if (!new_dev_maps) { 2840 mutex_unlock(&xps_map_mutex); 2841 return -ENOMEM; 2842 } 2843 2844 new_dev_maps->nr_ids = nr_ids; 2845 new_dev_maps->num_tc = num_tc; 2846 } 2847 2848 tci = j * num_tc + tc; 2849 map = copy ? xmap_dereference(dev_maps->attr_map[tci]) : NULL; 2850 2851 map = expand_xps_map(map, j, index, type == XPS_RXQS); 2852 if (!map) 2853 goto error; 2854 2855 RCU_INIT_POINTER(new_dev_maps->attr_map[tci], map); 2856 } 2857 2858 if (!new_dev_maps) 2859 goto out_no_new_maps; 2860 2861 if (!dev_maps) { 2862 /* Increment static keys at most once per type */ 2863 static_key_slow_inc_cpuslocked(&xps_needed); 2864 if (type == XPS_RXQS) 2865 static_key_slow_inc_cpuslocked(&xps_rxqs_needed); 2866 } 2867 2868 for (j = 0; j < nr_ids; j++) { 2869 bool skip_tc = false; 2870 2871 tci = j * num_tc + tc; 2872 if (netif_attr_test_mask(j, mask, nr_ids) && 2873 netif_attr_test_online(j, online_mask, nr_ids)) { 2874 /* add tx-queue to CPU/rx-queue maps */ 2875 int pos = 0; 2876 2877 skip_tc = true; 2878 2879 map = xmap_dereference(new_dev_maps->attr_map[tci]); 2880 while ((pos < map->len) && (map->queues[pos] != index)) 2881 pos++; 2882 2883 if (pos == map->len) 2884 map->queues[map->len++] = index; 2885 #ifdef CONFIG_NUMA 2886 if (type == XPS_CPUS) { 2887 if (numa_node_id == -2) 2888 numa_node_id = cpu_to_node(j); 2889 else if (numa_node_id != cpu_to_node(j)) 2890 numa_node_id = -1; 2891 } 2892 #endif 2893 } 2894 2895 if (copy) 2896 xps_copy_dev_maps(dev_maps, new_dev_maps, j, tc, 2897 skip_tc); 2898 } 2899 2900 rcu_assign_pointer(dev->xps_maps[type], new_dev_maps); 2901 2902 /* Cleanup old maps */ 2903 if (!dev_maps) 2904 goto out_no_old_maps; 2905 2906 for (j = 0; j < dev_maps->nr_ids; j++) { 2907 for (i = num_tc, tci = j * dev_maps->num_tc; i--; tci++) { 2908 map = xmap_dereference(dev_maps->attr_map[tci]); 2909 if (!map) 2910 continue; 2911 2912 if (copy) { 2913 new_map = xmap_dereference(new_dev_maps->attr_map[tci]); 2914 if (map == new_map) 2915 continue; 2916 } 2917 2918 RCU_INIT_POINTER(dev_maps->attr_map[tci], NULL); 2919 kfree_rcu(map, rcu); 2920 } 2921 } 2922 2923 old_dev_maps = dev_maps; 2924 2925 out_no_old_maps: 2926 dev_maps = new_dev_maps; 2927 active = true; 2928 2929 out_no_new_maps: 2930 if (type == XPS_CPUS) 2931 /* update Tx queue numa node */ 2932 netdev_queue_numa_node_write(netdev_get_tx_queue(dev, index), 2933 (numa_node_id >= 0) ? 2934 numa_node_id : NUMA_NO_NODE); 2935 2936 if (!dev_maps) 2937 goto out_no_maps; 2938 2939 /* removes tx-queue from unused CPUs/rx-queues */ 2940 for (j = 0; j < dev_maps->nr_ids; j++) { 2941 tci = j * dev_maps->num_tc; 2942 2943 for (i = 0; i < dev_maps->num_tc; i++, tci++) { 2944 if (i == tc && 2945 netif_attr_test_mask(j, mask, dev_maps->nr_ids) && 2946 netif_attr_test_online(j, online_mask, dev_maps->nr_ids)) 2947 continue; 2948 2949 active |= remove_xps_queue(dev_maps, 2950 copy ? old_dev_maps : NULL, 2951 tci, index); 2952 } 2953 } 2954 2955 if (old_dev_maps) 2956 kfree_rcu(old_dev_maps, rcu); 2957 2958 /* free map if not active */ 2959 if (!active) 2960 reset_xps_maps(dev, dev_maps, type); 2961 2962 out_no_maps: 2963 mutex_unlock(&xps_map_mutex); 2964 2965 return 0; 2966 error: 2967 /* remove any maps that we added */ 2968 for (j = 0; j < nr_ids; j++) { 2969 for (i = num_tc, tci = j * num_tc; i--; tci++) { 2970 new_map = xmap_dereference(new_dev_maps->attr_map[tci]); 2971 map = copy ? 2972 xmap_dereference(dev_maps->attr_map[tci]) : 2973 NULL; 2974 if (new_map && new_map != map) 2975 kfree(new_map); 2976 } 2977 } 2978 2979 mutex_unlock(&xps_map_mutex); 2980 2981 kfree(new_dev_maps); 2982 return -ENOMEM; 2983 } 2984 EXPORT_SYMBOL_GPL(__netif_set_xps_queue); 2985 2986 int netif_set_xps_queue(struct net_device *dev, const struct cpumask *mask, 2987 u16 index) 2988 { 2989 int ret; 2990 2991 cpus_read_lock(); 2992 ret = __netif_set_xps_queue(dev, cpumask_bits(mask), index, XPS_CPUS); 2993 cpus_read_unlock(); 2994 2995 return ret; 2996 } 2997 EXPORT_SYMBOL(netif_set_xps_queue); 2998 2999 #endif 3000 static void netdev_unbind_all_sb_channels(struct net_device *dev) 3001 { 3002 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues]; 3003 3004 /* Unbind any subordinate channels */ 3005 while (txq-- != &dev->_tx[0]) { 3006 if (txq->sb_dev) 3007 netdev_unbind_sb_channel(dev, txq->sb_dev); 3008 } 3009 } 3010 3011 void netdev_reset_tc(struct net_device *dev) 3012 { 3013 #ifdef CONFIG_XPS 3014 netif_reset_xps_queues_gt(dev, 0); 3015 #endif 3016 netdev_unbind_all_sb_channels(dev); 3017 3018 /* Reset TC configuration of device */ 3019 dev->num_tc = 0; 3020 memset(dev->tc_to_txq, 0, sizeof(dev->tc_to_txq)); 3021 memset(dev->prio_tc_map, 0, sizeof(dev->prio_tc_map)); 3022 } 3023 EXPORT_SYMBOL(netdev_reset_tc); 3024 3025 int netdev_set_tc_queue(struct net_device *dev, u8 tc, u16 count, u16 offset) 3026 { 3027 if (tc >= dev->num_tc) 3028 return -EINVAL; 3029 3030 #ifdef CONFIG_XPS 3031 netif_reset_xps_queues(dev, offset, count); 3032 #endif 3033 dev->tc_to_txq[tc].count = count; 3034 dev->tc_to_txq[tc].offset = offset; 3035 return 0; 3036 } 3037 EXPORT_SYMBOL(netdev_set_tc_queue); 3038 3039 int netdev_set_num_tc(struct net_device *dev, u8 num_tc) 3040 { 3041 if (num_tc > TC_MAX_QUEUE) 3042 return -EINVAL; 3043 3044 #ifdef CONFIG_XPS 3045 netif_reset_xps_queues_gt(dev, 0); 3046 #endif 3047 netdev_unbind_all_sb_channels(dev); 3048 3049 dev->num_tc = num_tc; 3050 return 0; 3051 } 3052 EXPORT_SYMBOL(netdev_set_num_tc); 3053 3054 void netdev_unbind_sb_channel(struct net_device *dev, 3055 struct net_device *sb_dev) 3056 { 3057 struct netdev_queue *txq = &dev->_tx[dev->num_tx_queues]; 3058 3059 #ifdef CONFIG_XPS 3060 netif_reset_xps_queues_gt(sb_dev, 0); 3061 #endif 3062 memset(sb_dev->tc_to_txq, 0, sizeof(sb_dev->tc_to_txq)); 3063 memset(sb_dev->prio_tc_map, 0, sizeof(sb_dev->prio_tc_map)); 3064 3065 while (txq-- != &dev->_tx[0]) { 3066 if (txq->sb_dev == sb_dev) 3067 txq->sb_dev = NULL; 3068 } 3069 } 3070 EXPORT_SYMBOL(netdev_unbind_sb_channel); 3071 3072 int netdev_bind_sb_channel_queue(struct net_device *dev, 3073 struct net_device *sb_dev, 3074 u8 tc, u16 count, u16 offset) 3075 { 3076 /* Make certain the sb_dev and dev are already configured */ 3077 if (sb_dev->num_tc >= 0 || tc >= dev->num_tc) 3078 return -EINVAL; 3079 3080 /* We cannot hand out queues we don't have */ 3081 if ((offset + count) > dev->real_num_tx_queues) 3082 return -EINVAL; 3083 3084 /* Record the mapping */ 3085 sb_dev->tc_to_txq[tc].count = count; 3086 sb_dev->tc_to_txq[tc].offset = offset; 3087 3088 /* Provide a way for Tx queue to find the tc_to_txq map or 3089 * XPS map for itself. 3090 */ 3091 while (count--) 3092 netdev_get_tx_queue(dev, count + offset)->sb_dev = sb_dev; 3093 3094 return 0; 3095 } 3096 EXPORT_SYMBOL(netdev_bind_sb_channel_queue); 3097 3098 int netdev_set_sb_channel(struct net_device *dev, u16 channel) 3099 { 3100 /* Do not use a multiqueue device to represent a subordinate channel */ 3101 if (netif_is_multiqueue(dev)) 3102 return -ENODEV; 3103 3104 /* We allow channels 1 - 32767 to be used for subordinate channels. 3105 * Channel 0 is meant to be "native" mode and used only to represent 3106 * the main root device. We allow writing 0 to reset the device back 3107 * to normal mode after being used as a subordinate channel. 3108 */ 3109 if (channel > S16_MAX) 3110 return -EINVAL; 3111 3112 dev->num_tc = -channel; 3113 3114 return 0; 3115 } 3116 EXPORT_SYMBOL(netdev_set_sb_channel); 3117 3118 /* 3119 * Routine to help set real_num_tx_queues. To avoid skbs mapped to queues 3120 * greater than real_num_tx_queues stale skbs on the qdisc must be flushed. 3121 */ 3122 int netif_set_real_num_tx_queues(struct net_device *dev, unsigned int txq) 3123 { 3124 bool disabling; 3125 int rc; 3126 3127 disabling = txq < dev->real_num_tx_queues; 3128 3129 if (txq < 1 || txq > dev->num_tx_queues) 3130 return -EINVAL; 3131 3132 if (dev->reg_state == NETREG_REGISTERED || 3133 dev->reg_state == NETREG_UNREGISTERING) { 3134 ASSERT_RTNL(); 3135 3136 rc = netdev_queue_update_kobjects(dev, dev->real_num_tx_queues, 3137 txq); 3138 if (rc) 3139 return rc; 3140 3141 if (dev->num_tc) 3142 netif_setup_tc(dev, txq); 3143 3144 net_shaper_set_real_num_tx_queues(dev, txq); 3145 3146 dev_qdisc_change_real_num_tx(dev, txq); 3147 3148 dev->real_num_tx_queues = txq; 3149 3150 if (disabling) { 3151 synchronize_net(); 3152 qdisc_reset_all_tx_gt(dev, txq); 3153 #ifdef CONFIG_XPS 3154 netif_reset_xps_queues_gt(dev, txq); 3155 #endif 3156 } 3157 } else { 3158 dev->real_num_tx_queues = txq; 3159 } 3160 3161 return 0; 3162 } 3163 EXPORT_SYMBOL(netif_set_real_num_tx_queues); 3164 3165 #ifdef CONFIG_SYSFS 3166 /** 3167 * netif_set_real_num_rx_queues - set actual number of RX queues used 3168 * @dev: Network device 3169 * @rxq: Actual number of RX queues 3170 * 3171 * This must be called either with the rtnl_lock held or before 3172 * registration of the net device. Returns 0 on success, or a 3173 * negative error code. If called before registration, it always 3174 * succeeds. 3175 */ 3176 int netif_set_real_num_rx_queues(struct net_device *dev, unsigned int rxq) 3177 { 3178 int rc; 3179 3180 if (rxq < 1 || rxq > dev->num_rx_queues) 3181 return -EINVAL; 3182 3183 if (dev->reg_state == NETREG_REGISTERED) { 3184 ASSERT_RTNL(); 3185 3186 rc = net_rx_queue_update_kobjects(dev, dev->real_num_rx_queues, 3187 rxq); 3188 if (rc) 3189 return rc; 3190 } 3191 3192 dev->real_num_rx_queues = rxq; 3193 return 0; 3194 } 3195 EXPORT_SYMBOL(netif_set_real_num_rx_queues); 3196 #endif 3197 3198 /** 3199 * netif_set_real_num_queues - set actual number of RX and TX queues used 3200 * @dev: Network device 3201 * @txq: Actual number of TX queues 3202 * @rxq: Actual number of RX queues 3203 * 3204 * Set the real number of both TX and RX queues. 3205 * Does nothing if the number of queues is already correct. 3206 */ 3207 int netif_set_real_num_queues(struct net_device *dev, 3208 unsigned int txq, unsigned int rxq) 3209 { 3210 unsigned int old_rxq = dev->real_num_rx_queues; 3211 int err; 3212 3213 if (txq < 1 || txq > dev->num_tx_queues || 3214 rxq < 1 || rxq > dev->num_rx_queues) 3215 return -EINVAL; 3216 3217 /* Start from increases, so the error path only does decreases - 3218 * decreases can't fail. 3219 */ 3220 if (rxq > dev->real_num_rx_queues) { 3221 err = netif_set_real_num_rx_queues(dev, rxq); 3222 if (err) 3223 return err; 3224 } 3225 if (txq > dev->real_num_tx_queues) { 3226 err = netif_set_real_num_tx_queues(dev, txq); 3227 if (err) 3228 goto undo_rx; 3229 } 3230 if (rxq < dev->real_num_rx_queues) 3231 WARN_ON(netif_set_real_num_rx_queues(dev, rxq)); 3232 if (txq < dev->real_num_tx_queues) 3233 WARN_ON(netif_set_real_num_tx_queues(dev, txq)); 3234 3235 return 0; 3236 undo_rx: 3237 WARN_ON(netif_set_real_num_rx_queues(dev, old_rxq)); 3238 return err; 3239 } 3240 EXPORT_SYMBOL(netif_set_real_num_queues); 3241 3242 /** 3243 * netif_set_tso_max_size() - set the max size of TSO frames supported 3244 * @dev: netdev to update 3245 * @size: max skb->len of a TSO frame 3246 * 3247 * Set the limit on the size of TSO super-frames the device can handle. 3248 * Unless explicitly set the stack will assume the value of 3249 * %GSO_LEGACY_MAX_SIZE. 3250 */ 3251 void netif_set_tso_max_size(struct net_device *dev, unsigned int size) 3252 { 3253 dev->tso_max_size = min(GSO_MAX_SIZE, size); 3254 if (size < READ_ONCE(dev->gso_max_size)) 3255 netif_set_gso_max_size(dev, size); 3256 if (size < READ_ONCE(dev->gso_ipv4_max_size)) 3257 netif_set_gso_ipv4_max_size(dev, size); 3258 } 3259 EXPORT_SYMBOL(netif_set_tso_max_size); 3260 3261 /** 3262 * netif_set_tso_max_segs() - set the max number of segs supported for TSO 3263 * @dev: netdev to update 3264 * @segs: max number of TCP segments 3265 * 3266 * Set the limit on the number of TCP segments the device can generate from 3267 * a single TSO super-frame. 3268 * Unless explicitly set the stack will assume the value of %GSO_MAX_SEGS. 3269 */ 3270 void netif_set_tso_max_segs(struct net_device *dev, unsigned int segs) 3271 { 3272 dev->tso_max_segs = segs; 3273 if (segs < READ_ONCE(dev->gso_max_segs)) 3274 netif_set_gso_max_segs(dev, segs); 3275 } 3276 EXPORT_SYMBOL(netif_set_tso_max_segs); 3277 3278 /** 3279 * netif_inherit_tso_max() - copy all TSO limits from a lower device to an upper 3280 * @to: netdev to update 3281 * @from: netdev from which to copy the limits 3282 */ 3283 void netif_inherit_tso_max(struct net_device *to, const struct net_device *from) 3284 { 3285 netif_set_tso_max_size(to, from->tso_max_size); 3286 netif_set_tso_max_segs(to, from->tso_max_segs); 3287 } 3288 EXPORT_SYMBOL(netif_inherit_tso_max); 3289 3290 /** 3291 * netif_get_num_default_rss_queues - default number of RSS queues 3292 * 3293 * Default value is the number of physical cores if there are only 1 or 2, or 3294 * divided by 2 if there are more. 3295 */ 3296 int netif_get_num_default_rss_queues(void) 3297 { 3298 cpumask_var_t cpus; 3299 int cpu, count = 0; 3300 3301 if (unlikely(is_kdump_kernel() || !zalloc_cpumask_var(&cpus, GFP_KERNEL))) 3302 return 1; 3303 3304 cpumask_copy(cpus, cpu_online_mask); 3305 for_each_cpu(cpu, cpus) { 3306 ++count; 3307 cpumask_andnot(cpus, cpus, topology_sibling_cpumask(cpu)); 3308 } 3309 free_cpumask_var(cpus); 3310 3311 return count > 2 ? DIV_ROUND_UP(count, 2) : count; 3312 } 3313 EXPORT_SYMBOL(netif_get_num_default_rss_queues); 3314 3315 static void __netif_reschedule(struct Qdisc *q) 3316 { 3317 struct softnet_data *sd; 3318 unsigned long flags; 3319 3320 local_irq_save(flags); 3321 sd = this_cpu_ptr(&softnet_data); 3322 q->next_sched = NULL; 3323 *sd->output_queue_tailp = q; 3324 sd->output_queue_tailp = &q->next_sched; 3325 raise_softirq_irqoff(NET_TX_SOFTIRQ); 3326 local_irq_restore(flags); 3327 } 3328 3329 void __netif_schedule(struct Qdisc *q) 3330 { 3331 if (!test_and_set_bit(__QDISC_STATE_SCHED, &q->state)) 3332 __netif_reschedule(q); 3333 } 3334 EXPORT_SYMBOL(__netif_schedule); 3335 3336 struct dev_kfree_skb_cb { 3337 enum skb_drop_reason reason; 3338 }; 3339 3340 static struct dev_kfree_skb_cb *get_kfree_skb_cb(const struct sk_buff *skb) 3341 { 3342 return (struct dev_kfree_skb_cb *)skb->cb; 3343 } 3344 3345 void netif_schedule_queue(struct netdev_queue *txq) 3346 { 3347 rcu_read_lock(); 3348 if (!netif_xmit_stopped(txq)) { 3349 struct Qdisc *q = rcu_dereference(txq->qdisc); 3350 3351 __netif_schedule(q); 3352 } 3353 rcu_read_unlock(); 3354 } 3355 EXPORT_SYMBOL(netif_schedule_queue); 3356 3357 void netif_tx_wake_queue(struct netdev_queue *dev_queue) 3358 { 3359 if (test_and_clear_bit(__QUEUE_STATE_DRV_XOFF, &dev_queue->state)) { 3360 struct Qdisc *q; 3361 3362 rcu_read_lock(); 3363 q = rcu_dereference(dev_queue->qdisc); 3364 __netif_schedule(q); 3365 rcu_read_unlock(); 3366 } 3367 } 3368 EXPORT_SYMBOL(netif_tx_wake_queue); 3369 3370 void dev_kfree_skb_irq_reason(struct sk_buff *skb, enum skb_drop_reason reason) 3371 { 3372 unsigned long flags; 3373 3374 if (unlikely(!skb)) 3375 return; 3376 3377 if (likely(refcount_read(&skb->users) == 1)) { 3378 smp_rmb(); 3379 refcount_set(&skb->users, 0); 3380 } else if (likely(!refcount_dec_and_test(&skb->users))) { 3381 return; 3382 } 3383 get_kfree_skb_cb(skb)->reason = reason; 3384 local_irq_save(flags); 3385 skb->next = __this_cpu_read(softnet_data.completion_queue); 3386 __this_cpu_write(softnet_data.completion_queue, skb); 3387 raise_softirq_irqoff(NET_TX_SOFTIRQ); 3388 local_irq_restore(flags); 3389 } 3390 EXPORT_SYMBOL(dev_kfree_skb_irq_reason); 3391 3392 void dev_kfree_skb_any_reason(struct sk_buff *skb, enum skb_drop_reason reason) 3393 { 3394 if (in_hardirq() || irqs_disabled()) 3395 dev_kfree_skb_irq_reason(skb, reason); 3396 else 3397 kfree_skb_reason(skb, reason); 3398 } 3399 EXPORT_SYMBOL(dev_kfree_skb_any_reason); 3400 3401 3402 /** 3403 * netif_device_detach - mark device as removed 3404 * @dev: network device 3405 * 3406 * Mark device as removed from system and therefore no longer available. 3407 */ 3408 void netif_device_detach(struct net_device *dev) 3409 { 3410 if (test_and_clear_bit(__LINK_STATE_PRESENT, &dev->state) && 3411 netif_running(dev)) { 3412 netif_tx_stop_all_queues(dev); 3413 } 3414 } 3415 EXPORT_SYMBOL(netif_device_detach); 3416 3417 /** 3418 * netif_device_attach - mark device as attached 3419 * @dev: network device 3420 * 3421 * Mark device as attached from system and restart if needed. 3422 */ 3423 void netif_device_attach(struct net_device *dev) 3424 { 3425 if (!test_and_set_bit(__LINK_STATE_PRESENT, &dev->state) && 3426 netif_running(dev)) { 3427 netif_tx_wake_all_queues(dev); 3428 netdev_watchdog_up(dev); 3429 } 3430 } 3431 EXPORT_SYMBOL(netif_device_attach); 3432 3433 /* 3434 * Returns a Tx hash based on the given packet descriptor a Tx queues' number 3435 * to be used as a distribution range. 3436 */ 3437 static u16 skb_tx_hash(const struct net_device *dev, 3438 const struct net_device *sb_dev, 3439 struct sk_buff *skb) 3440 { 3441 u32 hash; 3442 u16 qoffset = 0; 3443 u16 qcount = dev->real_num_tx_queues; 3444 3445 if (dev->num_tc) { 3446 u8 tc = netdev_get_prio_tc_map(dev, skb->priority); 3447 3448 qoffset = sb_dev->tc_to_txq[tc].offset; 3449 qcount = sb_dev->tc_to_txq[tc].count; 3450 if (unlikely(!qcount)) { 3451 net_warn_ratelimited("%s: invalid qcount, qoffset %u for tc %u\n", 3452 sb_dev->name, qoffset, tc); 3453 qoffset = 0; 3454 qcount = dev->real_num_tx_queues; 3455 } 3456 } 3457 3458 if (skb_rx_queue_recorded(skb)) { 3459 DEBUG_NET_WARN_ON_ONCE(qcount == 0); 3460 hash = skb_get_rx_queue(skb); 3461 if (hash >= qoffset) 3462 hash -= qoffset; 3463 while (unlikely(hash >= qcount)) 3464 hash -= qcount; 3465 return hash + qoffset; 3466 } 3467 3468 return (u16) reciprocal_scale(skb_get_hash(skb), qcount) + qoffset; 3469 } 3470 3471 void skb_warn_bad_offload(const struct sk_buff *skb) 3472 { 3473 static const netdev_features_t null_features; 3474 struct net_device *dev = skb->dev; 3475 const char *name = ""; 3476 3477 if (!net_ratelimit()) 3478 return; 3479 3480 if (dev) { 3481 if (dev->dev.parent) 3482 name = dev_driver_string(dev->dev.parent); 3483 else 3484 name = netdev_name(dev); 3485 } 3486 skb_dump(KERN_WARNING, skb, false); 3487 WARN(1, "%s: caps=(%pNF, %pNF)\n", 3488 name, dev ? &dev->features : &null_features, 3489 skb->sk ? &skb->sk->sk_route_caps : &null_features); 3490 } 3491 3492 /* 3493 * Invalidate hardware checksum when packet is to be mangled, and 3494 * complete checksum manually on outgoing path. 3495 */ 3496 int skb_checksum_help(struct sk_buff *skb) 3497 { 3498 __wsum csum; 3499 int ret = 0, offset; 3500 3501 if (skb->ip_summed == CHECKSUM_COMPLETE) 3502 goto out_set_summed; 3503 3504 if (unlikely(skb_is_gso(skb))) { 3505 skb_warn_bad_offload(skb); 3506 return -EINVAL; 3507 } 3508 3509 if (!skb_frags_readable(skb)) { 3510 return -EFAULT; 3511 } 3512 3513 /* Before computing a checksum, we should make sure no frag could 3514 * be modified by an external entity : checksum could be wrong. 3515 */ 3516 if (skb_has_shared_frag(skb)) { 3517 ret = __skb_linearize(skb); 3518 if (ret) 3519 goto out; 3520 } 3521 3522 offset = skb_checksum_start_offset(skb); 3523 ret = -EINVAL; 3524 if (unlikely(offset >= skb_headlen(skb))) { 3525 DO_ONCE_LITE(skb_dump, KERN_ERR, skb, false); 3526 WARN_ONCE(true, "offset (%d) >= skb_headlen() (%u)\n", 3527 offset, skb_headlen(skb)); 3528 goto out; 3529 } 3530 csum = skb_checksum(skb, offset, skb->len - offset, 0); 3531 3532 offset += skb->csum_offset; 3533 if (unlikely(offset + sizeof(__sum16) > skb_headlen(skb))) { 3534 DO_ONCE_LITE(skb_dump, KERN_ERR, skb, false); 3535 WARN_ONCE(true, "offset+2 (%zu) > skb_headlen() (%u)\n", 3536 offset + sizeof(__sum16), skb_headlen(skb)); 3537 goto out; 3538 } 3539 ret = skb_ensure_writable(skb, offset + sizeof(__sum16)); 3540 if (ret) 3541 goto out; 3542 3543 *(__sum16 *)(skb->data + offset) = csum_fold(csum) ?: CSUM_MANGLED_0; 3544 out_set_summed: 3545 skb->ip_summed = CHECKSUM_NONE; 3546 out: 3547 return ret; 3548 } 3549 EXPORT_SYMBOL(skb_checksum_help); 3550 3551 int skb_crc32c_csum_help(struct sk_buff *skb) 3552 { 3553 __le32 crc32c_csum; 3554 int ret = 0, offset, start; 3555 3556 if (skb->ip_summed != CHECKSUM_PARTIAL) 3557 goto out; 3558 3559 if (unlikely(skb_is_gso(skb))) 3560 goto out; 3561 3562 /* Before computing a checksum, we should make sure no frag could 3563 * be modified by an external entity : checksum could be wrong. 3564 */ 3565 if (unlikely(skb_has_shared_frag(skb))) { 3566 ret = __skb_linearize(skb); 3567 if (ret) 3568 goto out; 3569 } 3570 start = skb_checksum_start_offset(skb); 3571 offset = start + offsetof(struct sctphdr, checksum); 3572 if (WARN_ON_ONCE(offset >= skb_headlen(skb))) { 3573 ret = -EINVAL; 3574 goto out; 3575 } 3576 3577 ret = skb_ensure_writable(skb, offset + sizeof(__le32)); 3578 if (ret) 3579 goto out; 3580 3581 crc32c_csum = cpu_to_le32(~__skb_checksum(skb, start, 3582 skb->len - start, ~(__u32)0, 3583 crc32c_csum_stub)); 3584 *(__le32 *)(skb->data + offset) = crc32c_csum; 3585 skb_reset_csum_not_inet(skb); 3586 out: 3587 return ret; 3588 } 3589 EXPORT_SYMBOL(skb_crc32c_csum_help); 3590 3591 __be16 skb_network_protocol(struct sk_buff *skb, int *depth) 3592 { 3593 __be16 type = skb->protocol; 3594 3595 /* Tunnel gso handlers can set protocol to ethernet. */ 3596 if (type == htons(ETH_P_TEB)) { 3597 struct ethhdr *eth; 3598 3599 if (unlikely(!pskb_may_pull(skb, sizeof(struct ethhdr)))) 3600 return 0; 3601 3602 eth = (struct ethhdr *)skb->data; 3603 type = eth->h_proto; 3604 } 3605 3606 return vlan_get_protocol_and_depth(skb, type, depth); 3607 } 3608 3609 3610 /* Take action when hardware reception checksum errors are detected. */ 3611 #ifdef CONFIG_BUG 3612 static void do_netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb) 3613 { 3614 netdev_err(dev, "hw csum failure\n"); 3615 skb_dump(KERN_ERR, skb, true); 3616 dump_stack(); 3617 } 3618 3619 void netdev_rx_csum_fault(struct net_device *dev, struct sk_buff *skb) 3620 { 3621 DO_ONCE_LITE(do_netdev_rx_csum_fault, dev, skb); 3622 } 3623 EXPORT_SYMBOL(netdev_rx_csum_fault); 3624 #endif 3625 3626 /* XXX: check that highmem exists at all on the given machine. */ 3627 static int illegal_highdma(struct net_device *dev, struct sk_buff *skb) 3628 { 3629 #ifdef CONFIG_HIGHMEM 3630 int i; 3631 3632 if (!(dev->features & NETIF_F_HIGHDMA)) { 3633 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 3634 skb_frag_t *frag = &skb_shinfo(skb)->frags[i]; 3635 struct page *page = skb_frag_page(frag); 3636 3637 if (page && PageHighMem(page)) 3638 return 1; 3639 } 3640 } 3641 #endif 3642 return 0; 3643 } 3644 3645 /* If MPLS offload request, verify we are testing hardware MPLS features 3646 * instead of standard features for the netdev. 3647 */ 3648 #if IS_ENABLED(CONFIG_NET_MPLS_GSO) 3649 static netdev_features_t net_mpls_features(struct sk_buff *skb, 3650 netdev_features_t features, 3651 __be16 type) 3652 { 3653 if (eth_p_mpls(type)) 3654 features &= skb->dev->mpls_features; 3655 3656 return features; 3657 } 3658 #else 3659 static netdev_features_t net_mpls_features(struct sk_buff *skb, 3660 netdev_features_t features, 3661 __be16 type) 3662 { 3663 return features; 3664 } 3665 #endif 3666 3667 static netdev_features_t harmonize_features(struct sk_buff *skb, 3668 netdev_features_t features) 3669 { 3670 __be16 type; 3671 3672 type = skb_network_protocol(skb, NULL); 3673 features = net_mpls_features(skb, features, type); 3674 3675 if (skb->ip_summed != CHECKSUM_NONE && 3676 !can_checksum_protocol(features, type)) { 3677 features &= ~(NETIF_F_CSUM_MASK | NETIF_F_GSO_MASK); 3678 } 3679 if (illegal_highdma(skb->dev, skb)) 3680 features &= ~NETIF_F_SG; 3681 3682 return features; 3683 } 3684 3685 netdev_features_t passthru_features_check(struct sk_buff *skb, 3686 struct net_device *dev, 3687 netdev_features_t features) 3688 { 3689 return features; 3690 } 3691 EXPORT_SYMBOL(passthru_features_check); 3692 3693 static netdev_features_t dflt_features_check(struct sk_buff *skb, 3694 struct net_device *dev, 3695 netdev_features_t features) 3696 { 3697 return vlan_features_check(skb, features); 3698 } 3699 3700 static netdev_features_t gso_features_check(const struct sk_buff *skb, 3701 struct net_device *dev, 3702 netdev_features_t features) 3703 { 3704 u16 gso_segs = skb_shinfo(skb)->gso_segs; 3705 3706 if (gso_segs > READ_ONCE(dev->gso_max_segs)) 3707 return features & ~NETIF_F_GSO_MASK; 3708 3709 if (unlikely(skb->len >= netif_get_gso_max_size(dev, skb))) 3710 return features & ~NETIF_F_GSO_MASK; 3711 3712 if (!skb_shinfo(skb)->gso_type) { 3713 skb_warn_bad_offload(skb); 3714 return features & ~NETIF_F_GSO_MASK; 3715 } 3716 3717 /* Support for GSO partial features requires software 3718 * intervention before we can actually process the packets 3719 * so we need to strip support for any partial features now 3720 * and we can pull them back in after we have partially 3721 * segmented the frame. 3722 */ 3723 if (!(skb_shinfo(skb)->gso_type & SKB_GSO_PARTIAL)) 3724 features &= ~dev->gso_partial_features; 3725 3726 /* Make sure to clear the IPv4 ID mangling feature if the 3727 * IPv4 header has the potential to be fragmented. 3728 */ 3729 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4) { 3730 struct iphdr *iph = skb->encapsulation ? 3731 inner_ip_hdr(skb) : ip_hdr(skb); 3732 3733 if (!(iph->frag_off & htons(IP_DF))) 3734 features &= ~NETIF_F_TSO_MANGLEID; 3735 } 3736 3737 return features; 3738 } 3739 3740 netdev_features_t netif_skb_features(struct sk_buff *skb) 3741 { 3742 struct net_device *dev = skb->dev; 3743 netdev_features_t features = dev->features; 3744 3745 if (skb_is_gso(skb)) 3746 features = gso_features_check(skb, dev, features); 3747 3748 /* If encapsulation offload request, verify we are testing 3749 * hardware encapsulation features instead of standard 3750 * features for the netdev 3751 */ 3752 if (skb->encapsulation) 3753 features &= dev->hw_enc_features; 3754 3755 if (skb_vlan_tagged(skb)) 3756 features = netdev_intersect_features(features, 3757 dev->vlan_features | 3758 NETIF_F_HW_VLAN_CTAG_TX | 3759 NETIF_F_HW_VLAN_STAG_TX); 3760 3761 if (dev->netdev_ops->ndo_features_check) 3762 features &= dev->netdev_ops->ndo_features_check(skb, dev, 3763 features); 3764 else 3765 features &= dflt_features_check(skb, dev, features); 3766 3767 return harmonize_features(skb, features); 3768 } 3769 EXPORT_SYMBOL(netif_skb_features); 3770 3771 static int xmit_one(struct sk_buff *skb, struct net_device *dev, 3772 struct netdev_queue *txq, bool more) 3773 { 3774 unsigned int len; 3775 int rc; 3776 3777 if (dev_nit_active(dev)) 3778 dev_queue_xmit_nit(skb, dev); 3779 3780 len = skb->len; 3781 trace_net_dev_start_xmit(skb, dev); 3782 rc = netdev_start_xmit(skb, dev, txq, more); 3783 trace_net_dev_xmit(skb, rc, dev, len); 3784 3785 return rc; 3786 } 3787 3788 struct sk_buff *dev_hard_start_xmit(struct sk_buff *first, struct net_device *dev, 3789 struct netdev_queue *txq, int *ret) 3790 { 3791 struct sk_buff *skb = first; 3792 int rc = NETDEV_TX_OK; 3793 3794 while (skb) { 3795 struct sk_buff *next = skb->next; 3796 3797 skb_mark_not_on_list(skb); 3798 rc = xmit_one(skb, dev, txq, next != NULL); 3799 if (unlikely(!dev_xmit_complete(rc))) { 3800 skb->next = next; 3801 goto out; 3802 } 3803 3804 skb = next; 3805 if (netif_tx_queue_stopped(txq) && skb) { 3806 rc = NETDEV_TX_BUSY; 3807 break; 3808 } 3809 } 3810 3811 out: 3812 *ret = rc; 3813 return skb; 3814 } 3815 3816 static struct sk_buff *validate_xmit_vlan(struct sk_buff *skb, 3817 netdev_features_t features) 3818 { 3819 if (skb_vlan_tag_present(skb) && 3820 !vlan_hw_offload_capable(features, skb->vlan_proto)) 3821 skb = __vlan_hwaccel_push_inside(skb); 3822 return skb; 3823 } 3824 3825 int skb_csum_hwoffload_help(struct sk_buff *skb, 3826 const netdev_features_t features) 3827 { 3828 if (unlikely(skb_csum_is_sctp(skb))) 3829 return !!(features & NETIF_F_SCTP_CRC) ? 0 : 3830 skb_crc32c_csum_help(skb); 3831 3832 if (features & NETIF_F_HW_CSUM) 3833 return 0; 3834 3835 if (features & (NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM)) { 3836 if (vlan_get_protocol(skb) == htons(ETH_P_IPV6) && 3837 skb_network_header_len(skb) != sizeof(struct ipv6hdr) && 3838 !ipv6_has_hopopt_jumbo(skb)) 3839 goto sw_checksum; 3840 3841 switch (skb->csum_offset) { 3842 case offsetof(struct tcphdr, check): 3843 case offsetof(struct udphdr, check): 3844 return 0; 3845 } 3846 } 3847 3848 sw_checksum: 3849 return skb_checksum_help(skb); 3850 } 3851 EXPORT_SYMBOL(skb_csum_hwoffload_help); 3852 3853 static struct sk_buff *validate_xmit_skb(struct sk_buff *skb, struct net_device *dev, bool *again) 3854 { 3855 netdev_features_t features; 3856 3857 if (!skb_frags_readable(skb)) 3858 goto out_kfree_skb; 3859 3860 features = netif_skb_features(skb); 3861 skb = validate_xmit_vlan(skb, features); 3862 if (unlikely(!skb)) 3863 goto out_null; 3864 3865 skb = sk_validate_xmit_skb(skb, dev); 3866 if (unlikely(!skb)) 3867 goto out_null; 3868 3869 if (netif_needs_gso(skb, features)) { 3870 struct sk_buff *segs; 3871 3872 segs = skb_gso_segment(skb, features); 3873 if (IS_ERR(segs)) { 3874 goto out_kfree_skb; 3875 } else if (segs) { 3876 consume_skb(skb); 3877 skb = segs; 3878 } 3879 } else { 3880 if (skb_needs_linearize(skb, features) && 3881 __skb_linearize(skb)) 3882 goto out_kfree_skb; 3883 3884 /* If packet is not checksummed and device does not 3885 * support checksumming for this protocol, complete 3886 * checksumming here. 3887 */ 3888 if (skb->ip_summed == CHECKSUM_PARTIAL) { 3889 if (skb->encapsulation) 3890 skb_set_inner_transport_header(skb, 3891 skb_checksum_start_offset(skb)); 3892 else 3893 skb_set_transport_header(skb, 3894 skb_checksum_start_offset(skb)); 3895 if (skb_csum_hwoffload_help(skb, features)) 3896 goto out_kfree_skb; 3897 } 3898 } 3899 3900 skb = validate_xmit_xfrm(skb, features, again); 3901 3902 return skb; 3903 3904 out_kfree_skb: 3905 kfree_skb(skb); 3906 out_null: 3907 dev_core_stats_tx_dropped_inc(dev); 3908 return NULL; 3909 } 3910 3911 struct sk_buff *validate_xmit_skb_list(struct sk_buff *skb, struct net_device *dev, bool *again) 3912 { 3913 struct sk_buff *next, *head = NULL, *tail; 3914 3915 for (; skb != NULL; skb = next) { 3916 next = skb->next; 3917 skb_mark_not_on_list(skb); 3918 3919 /* in case skb won't be segmented, point to itself */ 3920 skb->prev = skb; 3921 3922 skb = validate_xmit_skb(skb, dev, again); 3923 if (!skb) 3924 continue; 3925 3926 if (!head) 3927 head = skb; 3928 else 3929 tail->next = skb; 3930 /* If skb was segmented, skb->prev points to 3931 * the last segment. If not, it still contains skb. 3932 */ 3933 tail = skb->prev; 3934 } 3935 return head; 3936 } 3937 EXPORT_SYMBOL_GPL(validate_xmit_skb_list); 3938 3939 static void qdisc_pkt_len_init(struct sk_buff *skb) 3940 { 3941 const struct skb_shared_info *shinfo = skb_shinfo(skb); 3942 3943 qdisc_skb_cb(skb)->pkt_len = skb->len; 3944 3945 /* To get more precise estimation of bytes sent on wire, 3946 * we add to pkt_len the headers size of all segments 3947 */ 3948 if (shinfo->gso_size && skb_transport_header_was_set(skb)) { 3949 u16 gso_segs = shinfo->gso_segs; 3950 unsigned int hdr_len; 3951 3952 /* mac layer + network layer */ 3953 hdr_len = skb_transport_offset(skb); 3954 3955 /* + transport layer */ 3956 if (likely(shinfo->gso_type & (SKB_GSO_TCPV4 | SKB_GSO_TCPV6))) { 3957 const struct tcphdr *th; 3958 struct tcphdr _tcphdr; 3959 3960 th = skb_header_pointer(skb, hdr_len, 3961 sizeof(_tcphdr), &_tcphdr); 3962 if (likely(th)) 3963 hdr_len += __tcp_hdrlen(th); 3964 } else if (shinfo->gso_type & SKB_GSO_UDP_L4) { 3965 struct udphdr _udphdr; 3966 3967 if (skb_header_pointer(skb, hdr_len, 3968 sizeof(_udphdr), &_udphdr)) 3969 hdr_len += sizeof(struct udphdr); 3970 } 3971 3972 if (unlikely(shinfo->gso_type & SKB_GSO_DODGY)) { 3973 int payload = skb->len - hdr_len; 3974 3975 /* Malicious packet. */ 3976 if (payload <= 0) 3977 return; 3978 gso_segs = DIV_ROUND_UP(payload, shinfo->gso_size); 3979 } 3980 qdisc_skb_cb(skb)->pkt_len += (gso_segs - 1) * hdr_len; 3981 } 3982 } 3983 3984 static int dev_qdisc_enqueue(struct sk_buff *skb, struct Qdisc *q, 3985 struct sk_buff **to_free, 3986 struct netdev_queue *txq) 3987 { 3988 int rc; 3989 3990 rc = q->enqueue(skb, q, to_free) & NET_XMIT_MASK; 3991 if (rc == NET_XMIT_SUCCESS) 3992 trace_qdisc_enqueue(q, txq, skb); 3993 return rc; 3994 } 3995 3996 static inline int __dev_xmit_skb(struct sk_buff *skb, struct Qdisc *q, 3997 struct net_device *dev, 3998 struct netdev_queue *txq) 3999 { 4000 spinlock_t *root_lock = qdisc_lock(q); 4001 struct sk_buff *to_free = NULL; 4002 bool contended; 4003 int rc; 4004 4005 qdisc_calculate_pkt_len(skb, q); 4006 4007 tcf_set_drop_reason(skb, SKB_DROP_REASON_QDISC_DROP); 4008 4009 if (q->flags & TCQ_F_NOLOCK) { 4010 if (q->flags & TCQ_F_CAN_BYPASS && nolock_qdisc_is_empty(q) && 4011 qdisc_run_begin(q)) { 4012 /* Retest nolock_qdisc_is_empty() within the protection 4013 * of q->seqlock to protect from racing with requeuing. 4014 */ 4015 if (unlikely(!nolock_qdisc_is_empty(q))) { 4016 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 4017 __qdisc_run(q); 4018 qdisc_run_end(q); 4019 4020 goto no_lock_out; 4021 } 4022 4023 qdisc_bstats_cpu_update(q, skb); 4024 if (sch_direct_xmit(skb, q, dev, txq, NULL, true) && 4025 !nolock_qdisc_is_empty(q)) 4026 __qdisc_run(q); 4027 4028 qdisc_run_end(q); 4029 return NET_XMIT_SUCCESS; 4030 } 4031 4032 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 4033 qdisc_run(q); 4034 4035 no_lock_out: 4036 if (unlikely(to_free)) 4037 kfree_skb_list_reason(to_free, 4038 tcf_get_drop_reason(to_free)); 4039 return rc; 4040 } 4041 4042 if (unlikely(READ_ONCE(q->owner) == smp_processor_id())) { 4043 kfree_skb_reason(skb, SKB_DROP_REASON_TC_RECLASSIFY_LOOP); 4044 return NET_XMIT_DROP; 4045 } 4046 /* 4047 * Heuristic to force contended enqueues to serialize on a 4048 * separate lock before trying to get qdisc main lock. 4049 * This permits qdisc->running owner to get the lock more 4050 * often and dequeue packets faster. 4051 * On PREEMPT_RT it is possible to preempt the qdisc owner during xmit 4052 * and then other tasks will only enqueue packets. The packets will be 4053 * sent after the qdisc owner is scheduled again. To prevent this 4054 * scenario the task always serialize on the lock. 4055 */ 4056 contended = qdisc_is_running(q) || IS_ENABLED(CONFIG_PREEMPT_RT); 4057 if (unlikely(contended)) 4058 spin_lock(&q->busylock); 4059 4060 spin_lock(root_lock); 4061 if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, &q->state))) { 4062 __qdisc_drop(skb, &to_free); 4063 rc = NET_XMIT_DROP; 4064 } else if ((q->flags & TCQ_F_CAN_BYPASS) && !qdisc_qlen(q) && 4065 qdisc_run_begin(q)) { 4066 /* 4067 * This is a work-conserving queue; there are no old skbs 4068 * waiting to be sent out; and the qdisc is not running - 4069 * xmit the skb directly. 4070 */ 4071 4072 qdisc_bstats_update(q, skb); 4073 4074 if (sch_direct_xmit(skb, q, dev, txq, root_lock, true)) { 4075 if (unlikely(contended)) { 4076 spin_unlock(&q->busylock); 4077 contended = false; 4078 } 4079 __qdisc_run(q); 4080 } 4081 4082 qdisc_run_end(q); 4083 rc = NET_XMIT_SUCCESS; 4084 } else { 4085 WRITE_ONCE(q->owner, smp_processor_id()); 4086 rc = dev_qdisc_enqueue(skb, q, &to_free, txq); 4087 WRITE_ONCE(q->owner, -1); 4088 if (qdisc_run_begin(q)) { 4089 if (unlikely(contended)) { 4090 spin_unlock(&q->busylock); 4091 contended = false; 4092 } 4093 __qdisc_run(q); 4094 qdisc_run_end(q); 4095 } 4096 } 4097 spin_unlock(root_lock); 4098 if (unlikely(to_free)) 4099 kfree_skb_list_reason(to_free, 4100 tcf_get_drop_reason(to_free)); 4101 if (unlikely(contended)) 4102 spin_unlock(&q->busylock); 4103 return rc; 4104 } 4105 4106 #if IS_ENABLED(CONFIG_CGROUP_NET_PRIO) 4107 static void skb_update_prio(struct sk_buff *skb) 4108 { 4109 const struct netprio_map *map; 4110 const struct sock *sk; 4111 unsigned int prioidx; 4112 4113 if (skb->priority) 4114 return; 4115 map = rcu_dereference_bh(skb->dev->priomap); 4116 if (!map) 4117 return; 4118 sk = skb_to_full_sk(skb); 4119 if (!sk) 4120 return; 4121 4122 prioidx = sock_cgroup_prioidx(&sk->sk_cgrp_data); 4123 4124 if (prioidx < map->priomap_len) 4125 skb->priority = map->priomap[prioidx]; 4126 } 4127 #else 4128 #define skb_update_prio(skb) 4129 #endif 4130 4131 /** 4132 * dev_loopback_xmit - loop back @skb 4133 * @net: network namespace this loopback is happening in 4134 * @sk: sk needed to be a netfilter okfn 4135 * @skb: buffer to transmit 4136 */ 4137 int dev_loopback_xmit(struct net *net, struct sock *sk, struct sk_buff *skb) 4138 { 4139 skb_reset_mac_header(skb); 4140 __skb_pull(skb, skb_network_offset(skb)); 4141 skb->pkt_type = PACKET_LOOPBACK; 4142 if (skb->ip_summed == CHECKSUM_NONE) 4143 skb->ip_summed = CHECKSUM_UNNECESSARY; 4144 DEBUG_NET_WARN_ON_ONCE(!skb_dst(skb)); 4145 skb_dst_force(skb); 4146 netif_rx(skb); 4147 return 0; 4148 } 4149 EXPORT_SYMBOL(dev_loopback_xmit); 4150 4151 #ifdef CONFIG_NET_EGRESS 4152 static struct netdev_queue * 4153 netdev_tx_queue_mapping(struct net_device *dev, struct sk_buff *skb) 4154 { 4155 int qm = skb_get_queue_mapping(skb); 4156 4157 return netdev_get_tx_queue(dev, netdev_cap_txqueue(dev, qm)); 4158 } 4159 4160 #ifndef CONFIG_PREEMPT_RT 4161 static bool netdev_xmit_txqueue_skipped(void) 4162 { 4163 return __this_cpu_read(softnet_data.xmit.skip_txqueue); 4164 } 4165 4166 void netdev_xmit_skip_txqueue(bool skip) 4167 { 4168 __this_cpu_write(softnet_data.xmit.skip_txqueue, skip); 4169 } 4170 EXPORT_SYMBOL_GPL(netdev_xmit_skip_txqueue); 4171 4172 #else 4173 static bool netdev_xmit_txqueue_skipped(void) 4174 { 4175 return current->net_xmit.skip_txqueue; 4176 } 4177 4178 void netdev_xmit_skip_txqueue(bool skip) 4179 { 4180 current->net_xmit.skip_txqueue = skip; 4181 } 4182 EXPORT_SYMBOL_GPL(netdev_xmit_skip_txqueue); 4183 #endif 4184 #endif /* CONFIG_NET_EGRESS */ 4185 4186 #ifdef CONFIG_NET_XGRESS 4187 static int tc_run(struct tcx_entry *entry, struct sk_buff *skb, 4188 enum skb_drop_reason *drop_reason) 4189 { 4190 int ret = TC_ACT_UNSPEC; 4191 #ifdef CONFIG_NET_CLS_ACT 4192 struct mini_Qdisc *miniq = rcu_dereference_bh(entry->miniq); 4193 struct tcf_result res; 4194 4195 if (!miniq) 4196 return ret; 4197 4198 /* Global bypass */ 4199 if (!static_branch_likely(&tcf_sw_enabled_key)) 4200 return ret; 4201 4202 /* Block-wise bypass */ 4203 if (tcf_block_bypass_sw(miniq->block)) 4204 return ret; 4205 4206 tc_skb_cb(skb)->mru = 0; 4207 tc_skb_cb(skb)->post_ct = false; 4208 tcf_set_drop_reason(skb, *drop_reason); 4209 4210 mini_qdisc_bstats_cpu_update(miniq, skb); 4211 ret = tcf_classify(skb, miniq->block, miniq->filter_list, &res, false); 4212 /* Only tcf related quirks below. */ 4213 switch (ret) { 4214 case TC_ACT_SHOT: 4215 *drop_reason = tcf_get_drop_reason(skb); 4216 mini_qdisc_qstats_cpu_drop(miniq); 4217 break; 4218 case TC_ACT_OK: 4219 case TC_ACT_RECLASSIFY: 4220 skb->tc_index = TC_H_MIN(res.classid); 4221 break; 4222 } 4223 #endif /* CONFIG_NET_CLS_ACT */ 4224 return ret; 4225 } 4226 4227 static DEFINE_STATIC_KEY_FALSE(tcx_needed_key); 4228 4229 void tcx_inc(void) 4230 { 4231 static_branch_inc(&tcx_needed_key); 4232 } 4233 4234 void tcx_dec(void) 4235 { 4236 static_branch_dec(&tcx_needed_key); 4237 } 4238 4239 static __always_inline enum tcx_action_base 4240 tcx_run(const struct bpf_mprog_entry *entry, struct sk_buff *skb, 4241 const bool needs_mac) 4242 { 4243 const struct bpf_mprog_fp *fp; 4244 const struct bpf_prog *prog; 4245 int ret = TCX_NEXT; 4246 4247 if (needs_mac) 4248 __skb_push(skb, skb->mac_len); 4249 bpf_mprog_foreach_prog(entry, fp, prog) { 4250 bpf_compute_data_pointers(skb); 4251 ret = bpf_prog_run(prog, skb); 4252 if (ret != TCX_NEXT) 4253 break; 4254 } 4255 if (needs_mac) 4256 __skb_pull(skb, skb->mac_len); 4257 return tcx_action_code(skb, ret); 4258 } 4259 4260 static __always_inline struct sk_buff * 4261 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret, 4262 struct net_device *orig_dev, bool *another) 4263 { 4264 struct bpf_mprog_entry *entry = rcu_dereference_bh(skb->dev->tcx_ingress); 4265 enum skb_drop_reason drop_reason = SKB_DROP_REASON_TC_INGRESS; 4266 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 4267 int sch_ret; 4268 4269 if (!entry) 4270 return skb; 4271 4272 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 4273 if (*pt_prev) { 4274 *ret = deliver_skb(skb, *pt_prev, orig_dev); 4275 *pt_prev = NULL; 4276 } 4277 4278 qdisc_skb_cb(skb)->pkt_len = skb->len; 4279 tcx_set_ingress(skb, true); 4280 4281 if (static_branch_unlikely(&tcx_needed_key)) { 4282 sch_ret = tcx_run(entry, skb, true); 4283 if (sch_ret != TC_ACT_UNSPEC) 4284 goto ingress_verdict; 4285 } 4286 sch_ret = tc_run(tcx_entry(entry), skb, &drop_reason); 4287 ingress_verdict: 4288 switch (sch_ret) { 4289 case TC_ACT_REDIRECT: 4290 /* skb_mac_header check was done by BPF, so we can safely 4291 * push the L2 header back before redirecting to another 4292 * netdev. 4293 */ 4294 __skb_push(skb, skb->mac_len); 4295 if (skb_do_redirect(skb) == -EAGAIN) { 4296 __skb_pull(skb, skb->mac_len); 4297 *another = true; 4298 break; 4299 } 4300 *ret = NET_RX_SUCCESS; 4301 bpf_net_ctx_clear(bpf_net_ctx); 4302 return NULL; 4303 case TC_ACT_SHOT: 4304 kfree_skb_reason(skb, drop_reason); 4305 *ret = NET_RX_DROP; 4306 bpf_net_ctx_clear(bpf_net_ctx); 4307 return NULL; 4308 /* used by tc_run */ 4309 case TC_ACT_STOLEN: 4310 case TC_ACT_QUEUED: 4311 case TC_ACT_TRAP: 4312 consume_skb(skb); 4313 fallthrough; 4314 case TC_ACT_CONSUMED: 4315 *ret = NET_RX_SUCCESS; 4316 bpf_net_ctx_clear(bpf_net_ctx); 4317 return NULL; 4318 } 4319 bpf_net_ctx_clear(bpf_net_ctx); 4320 4321 return skb; 4322 } 4323 4324 static __always_inline struct sk_buff * 4325 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev) 4326 { 4327 struct bpf_mprog_entry *entry = rcu_dereference_bh(dev->tcx_egress); 4328 enum skb_drop_reason drop_reason = SKB_DROP_REASON_TC_EGRESS; 4329 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 4330 int sch_ret; 4331 4332 if (!entry) 4333 return skb; 4334 4335 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 4336 4337 /* qdisc_skb_cb(skb)->pkt_len & tcx_set_ingress() was 4338 * already set by the caller. 4339 */ 4340 if (static_branch_unlikely(&tcx_needed_key)) { 4341 sch_ret = tcx_run(entry, skb, false); 4342 if (sch_ret != TC_ACT_UNSPEC) 4343 goto egress_verdict; 4344 } 4345 sch_ret = tc_run(tcx_entry(entry), skb, &drop_reason); 4346 egress_verdict: 4347 switch (sch_ret) { 4348 case TC_ACT_REDIRECT: 4349 /* No need to push/pop skb's mac_header here on egress! */ 4350 skb_do_redirect(skb); 4351 *ret = NET_XMIT_SUCCESS; 4352 bpf_net_ctx_clear(bpf_net_ctx); 4353 return NULL; 4354 case TC_ACT_SHOT: 4355 kfree_skb_reason(skb, drop_reason); 4356 *ret = NET_XMIT_DROP; 4357 bpf_net_ctx_clear(bpf_net_ctx); 4358 return NULL; 4359 /* used by tc_run */ 4360 case TC_ACT_STOLEN: 4361 case TC_ACT_QUEUED: 4362 case TC_ACT_TRAP: 4363 consume_skb(skb); 4364 fallthrough; 4365 case TC_ACT_CONSUMED: 4366 *ret = NET_XMIT_SUCCESS; 4367 bpf_net_ctx_clear(bpf_net_ctx); 4368 return NULL; 4369 } 4370 bpf_net_ctx_clear(bpf_net_ctx); 4371 4372 return skb; 4373 } 4374 #else 4375 static __always_inline struct sk_buff * 4376 sch_handle_ingress(struct sk_buff *skb, struct packet_type **pt_prev, int *ret, 4377 struct net_device *orig_dev, bool *another) 4378 { 4379 return skb; 4380 } 4381 4382 static __always_inline struct sk_buff * 4383 sch_handle_egress(struct sk_buff *skb, int *ret, struct net_device *dev) 4384 { 4385 return skb; 4386 } 4387 #endif /* CONFIG_NET_XGRESS */ 4388 4389 #ifdef CONFIG_XPS 4390 static int __get_xps_queue_idx(struct net_device *dev, struct sk_buff *skb, 4391 struct xps_dev_maps *dev_maps, unsigned int tci) 4392 { 4393 int tc = netdev_get_prio_tc_map(dev, skb->priority); 4394 struct xps_map *map; 4395 int queue_index = -1; 4396 4397 if (tc >= dev_maps->num_tc || tci >= dev_maps->nr_ids) 4398 return queue_index; 4399 4400 tci *= dev_maps->num_tc; 4401 tci += tc; 4402 4403 map = rcu_dereference(dev_maps->attr_map[tci]); 4404 if (map) { 4405 if (map->len == 1) 4406 queue_index = map->queues[0]; 4407 else 4408 queue_index = map->queues[reciprocal_scale( 4409 skb_get_hash(skb), map->len)]; 4410 if (unlikely(queue_index >= dev->real_num_tx_queues)) 4411 queue_index = -1; 4412 } 4413 return queue_index; 4414 } 4415 #endif 4416 4417 static int get_xps_queue(struct net_device *dev, struct net_device *sb_dev, 4418 struct sk_buff *skb) 4419 { 4420 #ifdef CONFIG_XPS 4421 struct xps_dev_maps *dev_maps; 4422 struct sock *sk = skb->sk; 4423 int queue_index = -1; 4424 4425 if (!static_key_false(&xps_needed)) 4426 return -1; 4427 4428 rcu_read_lock(); 4429 if (!static_key_false(&xps_rxqs_needed)) 4430 goto get_cpus_map; 4431 4432 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_RXQS]); 4433 if (dev_maps) { 4434 int tci = sk_rx_queue_get(sk); 4435 4436 if (tci >= 0) 4437 queue_index = __get_xps_queue_idx(dev, skb, dev_maps, 4438 tci); 4439 } 4440 4441 get_cpus_map: 4442 if (queue_index < 0) { 4443 dev_maps = rcu_dereference(sb_dev->xps_maps[XPS_CPUS]); 4444 if (dev_maps) { 4445 unsigned int tci = skb->sender_cpu - 1; 4446 4447 queue_index = __get_xps_queue_idx(dev, skb, dev_maps, 4448 tci); 4449 } 4450 } 4451 rcu_read_unlock(); 4452 4453 return queue_index; 4454 #else 4455 return -1; 4456 #endif 4457 } 4458 4459 u16 dev_pick_tx_zero(struct net_device *dev, struct sk_buff *skb, 4460 struct net_device *sb_dev) 4461 { 4462 return 0; 4463 } 4464 EXPORT_SYMBOL(dev_pick_tx_zero); 4465 4466 u16 netdev_pick_tx(struct net_device *dev, struct sk_buff *skb, 4467 struct net_device *sb_dev) 4468 { 4469 struct sock *sk = skb->sk; 4470 int queue_index = sk_tx_queue_get(sk); 4471 4472 sb_dev = sb_dev ? : dev; 4473 4474 if (queue_index < 0 || skb->ooo_okay || 4475 queue_index >= dev->real_num_tx_queues) { 4476 int new_index = get_xps_queue(dev, sb_dev, skb); 4477 4478 if (new_index < 0) 4479 new_index = skb_tx_hash(dev, sb_dev, skb); 4480 4481 if (queue_index != new_index && sk && 4482 sk_fullsock(sk) && 4483 rcu_access_pointer(sk->sk_dst_cache)) 4484 sk_tx_queue_set(sk, new_index); 4485 4486 queue_index = new_index; 4487 } 4488 4489 return queue_index; 4490 } 4491 EXPORT_SYMBOL(netdev_pick_tx); 4492 4493 struct netdev_queue *netdev_core_pick_tx(struct net_device *dev, 4494 struct sk_buff *skb, 4495 struct net_device *sb_dev) 4496 { 4497 int queue_index = 0; 4498 4499 #ifdef CONFIG_XPS 4500 u32 sender_cpu = skb->sender_cpu - 1; 4501 4502 if (sender_cpu >= (u32)NR_CPUS) 4503 skb->sender_cpu = raw_smp_processor_id() + 1; 4504 #endif 4505 4506 if (dev->real_num_tx_queues != 1) { 4507 const struct net_device_ops *ops = dev->netdev_ops; 4508 4509 if (ops->ndo_select_queue) 4510 queue_index = ops->ndo_select_queue(dev, skb, sb_dev); 4511 else 4512 queue_index = netdev_pick_tx(dev, skb, sb_dev); 4513 4514 queue_index = netdev_cap_txqueue(dev, queue_index); 4515 } 4516 4517 skb_set_queue_mapping(skb, queue_index); 4518 return netdev_get_tx_queue(dev, queue_index); 4519 } 4520 4521 /** 4522 * __dev_queue_xmit() - transmit a buffer 4523 * @skb: buffer to transmit 4524 * @sb_dev: suboordinate device used for L2 forwarding offload 4525 * 4526 * Queue a buffer for transmission to a network device. The caller must 4527 * have set the device and priority and built the buffer before calling 4528 * this function. The function can be called from an interrupt. 4529 * 4530 * When calling this method, interrupts MUST be enabled. This is because 4531 * the BH enable code must have IRQs enabled so that it will not deadlock. 4532 * 4533 * Regardless of the return value, the skb is consumed, so it is currently 4534 * difficult to retry a send to this method. (You can bump the ref count 4535 * before sending to hold a reference for retry if you are careful.) 4536 * 4537 * Return: 4538 * * 0 - buffer successfully transmitted 4539 * * positive qdisc return code - NET_XMIT_DROP etc. 4540 * * negative errno - other errors 4541 */ 4542 int __dev_queue_xmit(struct sk_buff *skb, struct net_device *sb_dev) 4543 { 4544 struct net_device *dev = skb->dev; 4545 struct netdev_queue *txq = NULL; 4546 struct Qdisc *q; 4547 int rc = -ENOMEM; 4548 bool again = false; 4549 4550 skb_reset_mac_header(skb); 4551 skb_assert_len(skb); 4552 4553 if (unlikely(skb_shinfo(skb)->tx_flags & 4554 (SKBTX_SCHED_TSTAMP | SKBTX_BPF))) 4555 __skb_tstamp_tx(skb, NULL, NULL, skb->sk, SCM_TSTAMP_SCHED); 4556 4557 /* Disable soft irqs for various locks below. Also 4558 * stops preemption for RCU. 4559 */ 4560 rcu_read_lock_bh(); 4561 4562 skb_update_prio(skb); 4563 4564 qdisc_pkt_len_init(skb); 4565 tcx_set_ingress(skb, false); 4566 #ifdef CONFIG_NET_EGRESS 4567 if (static_branch_unlikely(&egress_needed_key)) { 4568 if (nf_hook_egress_active()) { 4569 skb = nf_hook_egress(skb, &rc, dev); 4570 if (!skb) 4571 goto out; 4572 } 4573 4574 netdev_xmit_skip_txqueue(false); 4575 4576 nf_skip_egress(skb, true); 4577 skb = sch_handle_egress(skb, &rc, dev); 4578 if (!skb) 4579 goto out; 4580 nf_skip_egress(skb, false); 4581 4582 if (netdev_xmit_txqueue_skipped()) 4583 txq = netdev_tx_queue_mapping(dev, skb); 4584 } 4585 #endif 4586 /* If device/qdisc don't need skb->dst, release it right now while 4587 * its hot in this cpu cache. 4588 */ 4589 if (dev->priv_flags & IFF_XMIT_DST_RELEASE) 4590 skb_dst_drop(skb); 4591 else 4592 skb_dst_force(skb); 4593 4594 if (!txq) 4595 txq = netdev_core_pick_tx(dev, skb, sb_dev); 4596 4597 q = rcu_dereference_bh(txq->qdisc); 4598 4599 trace_net_dev_queue(skb); 4600 if (q->enqueue) { 4601 rc = __dev_xmit_skb(skb, q, dev, txq); 4602 goto out; 4603 } 4604 4605 /* The device has no queue. Common case for software devices: 4606 * loopback, all the sorts of tunnels... 4607 4608 * Really, it is unlikely that netif_tx_lock protection is necessary 4609 * here. (f.e. loopback and IP tunnels are clean ignoring statistics 4610 * counters.) 4611 * However, it is possible, that they rely on protection 4612 * made by us here. 4613 4614 * Check this and shot the lock. It is not prone from deadlocks. 4615 *Either shot noqueue qdisc, it is even simpler 8) 4616 */ 4617 if (dev->flags & IFF_UP) { 4618 int cpu = smp_processor_id(); /* ok because BHs are off */ 4619 4620 /* Other cpus might concurrently change txq->xmit_lock_owner 4621 * to -1 or to their cpu id, but not to our id. 4622 */ 4623 if (READ_ONCE(txq->xmit_lock_owner) != cpu) { 4624 if (dev_xmit_recursion()) 4625 goto recursion_alert; 4626 4627 skb = validate_xmit_skb(skb, dev, &again); 4628 if (!skb) 4629 goto out; 4630 4631 HARD_TX_LOCK(dev, txq, cpu); 4632 4633 if (!netif_xmit_stopped(txq)) { 4634 dev_xmit_recursion_inc(); 4635 skb = dev_hard_start_xmit(skb, dev, txq, &rc); 4636 dev_xmit_recursion_dec(); 4637 if (dev_xmit_complete(rc)) { 4638 HARD_TX_UNLOCK(dev, txq); 4639 goto out; 4640 } 4641 } 4642 HARD_TX_UNLOCK(dev, txq); 4643 net_crit_ratelimited("Virtual device %s asks to queue packet!\n", 4644 dev->name); 4645 } else { 4646 /* Recursion is detected! It is possible, 4647 * unfortunately 4648 */ 4649 recursion_alert: 4650 net_crit_ratelimited("Dead loop on virtual device %s, fix it urgently!\n", 4651 dev->name); 4652 } 4653 } 4654 4655 rc = -ENETDOWN; 4656 rcu_read_unlock_bh(); 4657 4658 dev_core_stats_tx_dropped_inc(dev); 4659 kfree_skb_list(skb); 4660 return rc; 4661 out: 4662 rcu_read_unlock_bh(); 4663 return rc; 4664 } 4665 EXPORT_SYMBOL(__dev_queue_xmit); 4666 4667 int __dev_direct_xmit(struct sk_buff *skb, u16 queue_id) 4668 { 4669 struct net_device *dev = skb->dev; 4670 struct sk_buff *orig_skb = skb; 4671 struct netdev_queue *txq; 4672 int ret = NETDEV_TX_BUSY; 4673 bool again = false; 4674 4675 if (unlikely(!netif_running(dev) || 4676 !netif_carrier_ok(dev))) 4677 goto drop; 4678 4679 skb = validate_xmit_skb_list(skb, dev, &again); 4680 if (skb != orig_skb) 4681 goto drop; 4682 4683 skb_set_queue_mapping(skb, queue_id); 4684 txq = skb_get_tx_queue(dev, skb); 4685 4686 local_bh_disable(); 4687 4688 dev_xmit_recursion_inc(); 4689 HARD_TX_LOCK(dev, txq, smp_processor_id()); 4690 if (!netif_xmit_frozen_or_drv_stopped(txq)) 4691 ret = netdev_start_xmit(skb, dev, txq, false); 4692 HARD_TX_UNLOCK(dev, txq); 4693 dev_xmit_recursion_dec(); 4694 4695 local_bh_enable(); 4696 return ret; 4697 drop: 4698 dev_core_stats_tx_dropped_inc(dev); 4699 kfree_skb_list(skb); 4700 return NET_XMIT_DROP; 4701 } 4702 EXPORT_SYMBOL(__dev_direct_xmit); 4703 4704 /************************************************************************* 4705 * Receiver routines 4706 *************************************************************************/ 4707 static DEFINE_PER_CPU(struct task_struct *, backlog_napi); 4708 4709 int weight_p __read_mostly = 64; /* old backlog weight */ 4710 int dev_weight_rx_bias __read_mostly = 1; /* bias for backlog weight */ 4711 int dev_weight_tx_bias __read_mostly = 1; /* bias for output_queue quota */ 4712 4713 /* Called with irq disabled */ 4714 static inline void ____napi_schedule(struct softnet_data *sd, 4715 struct napi_struct *napi) 4716 { 4717 struct task_struct *thread; 4718 4719 lockdep_assert_irqs_disabled(); 4720 4721 if (test_bit(NAPI_STATE_THREADED, &napi->state)) { 4722 /* Paired with smp_mb__before_atomic() in 4723 * napi_enable()/dev_set_threaded(). 4724 * Use READ_ONCE() to guarantee a complete 4725 * read on napi->thread. Only call 4726 * wake_up_process() when it's not NULL. 4727 */ 4728 thread = READ_ONCE(napi->thread); 4729 if (thread) { 4730 if (use_backlog_threads() && thread == raw_cpu_read(backlog_napi)) 4731 goto use_local_napi; 4732 4733 set_bit(NAPI_STATE_SCHED_THREADED, &napi->state); 4734 wake_up_process(thread); 4735 return; 4736 } 4737 } 4738 4739 use_local_napi: 4740 list_add_tail(&napi->poll_list, &sd->poll_list); 4741 WRITE_ONCE(napi->list_owner, smp_processor_id()); 4742 /* If not called from net_rx_action() 4743 * we have to raise NET_RX_SOFTIRQ. 4744 */ 4745 if (!sd->in_net_rx_action) 4746 raise_softirq_irqoff(NET_RX_SOFTIRQ); 4747 } 4748 4749 #ifdef CONFIG_RPS 4750 4751 struct static_key_false rps_needed __read_mostly; 4752 EXPORT_SYMBOL(rps_needed); 4753 struct static_key_false rfs_needed __read_mostly; 4754 EXPORT_SYMBOL(rfs_needed); 4755 4756 static struct rps_dev_flow * 4757 set_rps_cpu(struct net_device *dev, struct sk_buff *skb, 4758 struct rps_dev_flow *rflow, u16 next_cpu) 4759 { 4760 if (next_cpu < nr_cpu_ids) { 4761 u32 head; 4762 #ifdef CONFIG_RFS_ACCEL 4763 struct netdev_rx_queue *rxqueue; 4764 struct rps_dev_flow_table *flow_table; 4765 struct rps_dev_flow *old_rflow; 4766 u16 rxq_index; 4767 u32 flow_id; 4768 int rc; 4769 4770 /* Should we steer this flow to a different hardware queue? */ 4771 if (!skb_rx_queue_recorded(skb) || !dev->rx_cpu_rmap || 4772 !(dev->features & NETIF_F_NTUPLE)) 4773 goto out; 4774 rxq_index = cpu_rmap_lookup_index(dev->rx_cpu_rmap, next_cpu); 4775 if (rxq_index == skb_get_rx_queue(skb)) 4776 goto out; 4777 4778 rxqueue = dev->_rx + rxq_index; 4779 flow_table = rcu_dereference(rxqueue->rps_flow_table); 4780 if (!flow_table) 4781 goto out; 4782 flow_id = skb_get_hash(skb) & flow_table->mask; 4783 rc = dev->netdev_ops->ndo_rx_flow_steer(dev, skb, 4784 rxq_index, flow_id); 4785 if (rc < 0) 4786 goto out; 4787 old_rflow = rflow; 4788 rflow = &flow_table->flows[flow_id]; 4789 WRITE_ONCE(rflow->filter, rc); 4790 if (old_rflow->filter == rc) 4791 WRITE_ONCE(old_rflow->filter, RPS_NO_FILTER); 4792 out: 4793 #endif 4794 head = READ_ONCE(per_cpu(softnet_data, next_cpu).input_queue_head); 4795 rps_input_queue_tail_save(&rflow->last_qtail, head); 4796 } 4797 4798 WRITE_ONCE(rflow->cpu, next_cpu); 4799 return rflow; 4800 } 4801 4802 /* 4803 * get_rps_cpu is called from netif_receive_skb and returns the target 4804 * CPU from the RPS map of the receiving queue for a given skb. 4805 * rcu_read_lock must be held on entry. 4806 */ 4807 static int get_rps_cpu(struct net_device *dev, struct sk_buff *skb, 4808 struct rps_dev_flow **rflowp) 4809 { 4810 const struct rps_sock_flow_table *sock_flow_table; 4811 struct netdev_rx_queue *rxqueue = dev->_rx; 4812 struct rps_dev_flow_table *flow_table; 4813 struct rps_map *map; 4814 int cpu = -1; 4815 u32 tcpu; 4816 u32 hash; 4817 4818 if (skb_rx_queue_recorded(skb)) { 4819 u16 index = skb_get_rx_queue(skb); 4820 4821 if (unlikely(index >= dev->real_num_rx_queues)) { 4822 WARN_ONCE(dev->real_num_rx_queues > 1, 4823 "%s received packet on queue %u, but number " 4824 "of RX queues is %u\n", 4825 dev->name, index, dev->real_num_rx_queues); 4826 goto done; 4827 } 4828 rxqueue += index; 4829 } 4830 4831 /* Avoid computing hash if RFS/RPS is not active for this rxqueue */ 4832 4833 flow_table = rcu_dereference(rxqueue->rps_flow_table); 4834 map = rcu_dereference(rxqueue->rps_map); 4835 if (!flow_table && !map) 4836 goto done; 4837 4838 skb_reset_network_header(skb); 4839 hash = skb_get_hash(skb); 4840 if (!hash) 4841 goto done; 4842 4843 sock_flow_table = rcu_dereference(net_hotdata.rps_sock_flow_table); 4844 if (flow_table && sock_flow_table) { 4845 struct rps_dev_flow *rflow; 4846 u32 next_cpu; 4847 u32 ident; 4848 4849 /* First check into global flow table if there is a match. 4850 * This READ_ONCE() pairs with WRITE_ONCE() from rps_record_sock_flow(). 4851 */ 4852 ident = READ_ONCE(sock_flow_table->ents[hash & sock_flow_table->mask]); 4853 if ((ident ^ hash) & ~net_hotdata.rps_cpu_mask) 4854 goto try_rps; 4855 4856 next_cpu = ident & net_hotdata.rps_cpu_mask; 4857 4858 /* OK, now we know there is a match, 4859 * we can look at the local (per receive queue) flow table 4860 */ 4861 rflow = &flow_table->flows[hash & flow_table->mask]; 4862 tcpu = rflow->cpu; 4863 4864 /* 4865 * If the desired CPU (where last recvmsg was done) is 4866 * different from current CPU (one in the rx-queue flow 4867 * table entry), switch if one of the following holds: 4868 * - Current CPU is unset (>= nr_cpu_ids). 4869 * - Current CPU is offline. 4870 * - The current CPU's queue tail has advanced beyond the 4871 * last packet that was enqueued using this table entry. 4872 * This guarantees that all previous packets for the flow 4873 * have been dequeued, thus preserving in order delivery. 4874 */ 4875 if (unlikely(tcpu != next_cpu) && 4876 (tcpu >= nr_cpu_ids || !cpu_online(tcpu) || 4877 ((int)(READ_ONCE(per_cpu(softnet_data, tcpu).input_queue_head) - 4878 rflow->last_qtail)) >= 0)) { 4879 tcpu = next_cpu; 4880 rflow = set_rps_cpu(dev, skb, rflow, next_cpu); 4881 } 4882 4883 if (tcpu < nr_cpu_ids && cpu_online(tcpu)) { 4884 *rflowp = rflow; 4885 cpu = tcpu; 4886 goto done; 4887 } 4888 } 4889 4890 try_rps: 4891 4892 if (map) { 4893 tcpu = map->cpus[reciprocal_scale(hash, map->len)]; 4894 if (cpu_online(tcpu)) { 4895 cpu = tcpu; 4896 goto done; 4897 } 4898 } 4899 4900 done: 4901 return cpu; 4902 } 4903 4904 #ifdef CONFIG_RFS_ACCEL 4905 4906 /** 4907 * rps_may_expire_flow - check whether an RFS hardware filter may be removed 4908 * @dev: Device on which the filter was set 4909 * @rxq_index: RX queue index 4910 * @flow_id: Flow ID passed to ndo_rx_flow_steer() 4911 * @filter_id: Filter ID returned by ndo_rx_flow_steer() 4912 * 4913 * Drivers that implement ndo_rx_flow_steer() should periodically call 4914 * this function for each installed filter and remove the filters for 4915 * which it returns %true. 4916 */ 4917 bool rps_may_expire_flow(struct net_device *dev, u16 rxq_index, 4918 u32 flow_id, u16 filter_id) 4919 { 4920 struct netdev_rx_queue *rxqueue = dev->_rx + rxq_index; 4921 struct rps_dev_flow_table *flow_table; 4922 struct rps_dev_flow *rflow; 4923 bool expire = true; 4924 unsigned int cpu; 4925 4926 rcu_read_lock(); 4927 flow_table = rcu_dereference(rxqueue->rps_flow_table); 4928 if (flow_table && flow_id <= flow_table->mask) { 4929 rflow = &flow_table->flows[flow_id]; 4930 cpu = READ_ONCE(rflow->cpu); 4931 if (READ_ONCE(rflow->filter) == filter_id && cpu < nr_cpu_ids && 4932 ((int)(READ_ONCE(per_cpu(softnet_data, cpu).input_queue_head) - 4933 READ_ONCE(rflow->last_qtail)) < 4934 (int)(10 * flow_table->mask))) 4935 expire = false; 4936 } 4937 rcu_read_unlock(); 4938 return expire; 4939 } 4940 EXPORT_SYMBOL(rps_may_expire_flow); 4941 4942 #endif /* CONFIG_RFS_ACCEL */ 4943 4944 /* Called from hardirq (IPI) context */ 4945 static void rps_trigger_softirq(void *data) 4946 { 4947 struct softnet_data *sd = data; 4948 4949 ____napi_schedule(sd, &sd->backlog); 4950 sd->received_rps++; 4951 } 4952 4953 #endif /* CONFIG_RPS */ 4954 4955 /* Called from hardirq (IPI) context */ 4956 static void trigger_rx_softirq(void *data) 4957 { 4958 struct softnet_data *sd = data; 4959 4960 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 4961 smp_store_release(&sd->defer_ipi_scheduled, 0); 4962 } 4963 4964 /* 4965 * After we queued a packet into sd->input_pkt_queue, 4966 * we need to make sure this queue is serviced soon. 4967 * 4968 * - If this is another cpu queue, link it to our rps_ipi_list, 4969 * and make sure we will process rps_ipi_list from net_rx_action(). 4970 * 4971 * - If this is our own queue, NAPI schedule our backlog. 4972 * Note that this also raises NET_RX_SOFTIRQ. 4973 */ 4974 static void napi_schedule_rps(struct softnet_data *sd) 4975 { 4976 struct softnet_data *mysd = this_cpu_ptr(&softnet_data); 4977 4978 #ifdef CONFIG_RPS 4979 if (sd != mysd) { 4980 if (use_backlog_threads()) { 4981 __napi_schedule_irqoff(&sd->backlog); 4982 return; 4983 } 4984 4985 sd->rps_ipi_next = mysd->rps_ipi_list; 4986 mysd->rps_ipi_list = sd; 4987 4988 /* If not called from net_rx_action() or napi_threaded_poll() 4989 * we have to raise NET_RX_SOFTIRQ. 4990 */ 4991 if (!mysd->in_net_rx_action && !mysd->in_napi_threaded_poll) 4992 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 4993 return; 4994 } 4995 #endif /* CONFIG_RPS */ 4996 __napi_schedule_irqoff(&mysd->backlog); 4997 } 4998 4999 void kick_defer_list_purge(struct softnet_data *sd, unsigned int cpu) 5000 { 5001 unsigned long flags; 5002 5003 if (use_backlog_threads()) { 5004 backlog_lock_irq_save(sd, &flags); 5005 5006 if (!__test_and_set_bit(NAPI_STATE_SCHED, &sd->backlog.state)) 5007 __napi_schedule_irqoff(&sd->backlog); 5008 5009 backlog_unlock_irq_restore(sd, &flags); 5010 5011 } else if (!cmpxchg(&sd->defer_ipi_scheduled, 0, 1)) { 5012 smp_call_function_single_async(cpu, &sd->defer_csd); 5013 } 5014 } 5015 5016 #ifdef CONFIG_NET_FLOW_LIMIT 5017 int netdev_flow_limit_table_len __read_mostly = (1 << 12); 5018 #endif 5019 5020 static bool skb_flow_limit(struct sk_buff *skb, unsigned int qlen) 5021 { 5022 #ifdef CONFIG_NET_FLOW_LIMIT 5023 struct sd_flow_limit *fl; 5024 struct softnet_data *sd; 5025 unsigned int old_flow, new_flow; 5026 5027 if (qlen < (READ_ONCE(net_hotdata.max_backlog) >> 1)) 5028 return false; 5029 5030 sd = this_cpu_ptr(&softnet_data); 5031 5032 rcu_read_lock(); 5033 fl = rcu_dereference(sd->flow_limit); 5034 if (fl) { 5035 new_flow = skb_get_hash(skb) & (fl->num_buckets - 1); 5036 old_flow = fl->history[fl->history_head]; 5037 fl->history[fl->history_head] = new_flow; 5038 5039 fl->history_head++; 5040 fl->history_head &= FLOW_LIMIT_HISTORY - 1; 5041 5042 if (likely(fl->buckets[old_flow])) 5043 fl->buckets[old_flow]--; 5044 5045 if (++fl->buckets[new_flow] > (FLOW_LIMIT_HISTORY >> 1)) { 5046 fl->count++; 5047 rcu_read_unlock(); 5048 return true; 5049 } 5050 } 5051 rcu_read_unlock(); 5052 #endif 5053 return false; 5054 } 5055 5056 /* 5057 * enqueue_to_backlog is called to queue an skb to a per CPU backlog 5058 * queue (may be a remote CPU queue). 5059 */ 5060 static int enqueue_to_backlog(struct sk_buff *skb, int cpu, 5061 unsigned int *qtail) 5062 { 5063 enum skb_drop_reason reason; 5064 struct softnet_data *sd; 5065 unsigned long flags; 5066 unsigned int qlen; 5067 int max_backlog; 5068 u32 tail; 5069 5070 reason = SKB_DROP_REASON_DEV_READY; 5071 if (!netif_running(skb->dev)) 5072 goto bad_dev; 5073 5074 reason = SKB_DROP_REASON_CPU_BACKLOG; 5075 sd = &per_cpu(softnet_data, cpu); 5076 5077 qlen = skb_queue_len_lockless(&sd->input_pkt_queue); 5078 max_backlog = READ_ONCE(net_hotdata.max_backlog); 5079 if (unlikely(qlen > max_backlog)) 5080 goto cpu_backlog_drop; 5081 backlog_lock_irq_save(sd, &flags); 5082 qlen = skb_queue_len(&sd->input_pkt_queue); 5083 if (qlen <= max_backlog && !skb_flow_limit(skb, qlen)) { 5084 if (!qlen) { 5085 /* Schedule NAPI for backlog device. We can use 5086 * non atomic operation as we own the queue lock. 5087 */ 5088 if (!__test_and_set_bit(NAPI_STATE_SCHED, 5089 &sd->backlog.state)) 5090 napi_schedule_rps(sd); 5091 } 5092 __skb_queue_tail(&sd->input_pkt_queue, skb); 5093 tail = rps_input_queue_tail_incr(sd); 5094 backlog_unlock_irq_restore(sd, &flags); 5095 5096 /* save the tail outside of the critical section */ 5097 rps_input_queue_tail_save(qtail, tail); 5098 return NET_RX_SUCCESS; 5099 } 5100 5101 backlog_unlock_irq_restore(sd, &flags); 5102 5103 cpu_backlog_drop: 5104 atomic_inc(&sd->dropped); 5105 bad_dev: 5106 dev_core_stats_rx_dropped_inc(skb->dev); 5107 kfree_skb_reason(skb, reason); 5108 return NET_RX_DROP; 5109 } 5110 5111 static struct netdev_rx_queue *netif_get_rxqueue(struct sk_buff *skb) 5112 { 5113 struct net_device *dev = skb->dev; 5114 struct netdev_rx_queue *rxqueue; 5115 5116 rxqueue = dev->_rx; 5117 5118 if (skb_rx_queue_recorded(skb)) { 5119 u16 index = skb_get_rx_queue(skb); 5120 5121 if (unlikely(index >= dev->real_num_rx_queues)) { 5122 WARN_ONCE(dev->real_num_rx_queues > 1, 5123 "%s received packet on queue %u, but number " 5124 "of RX queues is %u\n", 5125 dev->name, index, dev->real_num_rx_queues); 5126 5127 return rxqueue; /* Return first rxqueue */ 5128 } 5129 rxqueue += index; 5130 } 5131 return rxqueue; 5132 } 5133 5134 u32 bpf_prog_run_generic_xdp(struct sk_buff *skb, struct xdp_buff *xdp, 5135 const struct bpf_prog *xdp_prog) 5136 { 5137 void *orig_data, *orig_data_end, *hard_start; 5138 struct netdev_rx_queue *rxqueue; 5139 bool orig_bcast, orig_host; 5140 u32 mac_len, frame_sz; 5141 __be16 orig_eth_type; 5142 struct ethhdr *eth; 5143 u32 metalen, act; 5144 int off; 5145 5146 /* The XDP program wants to see the packet starting at the MAC 5147 * header. 5148 */ 5149 mac_len = skb->data - skb_mac_header(skb); 5150 hard_start = skb->data - skb_headroom(skb); 5151 5152 /* SKB "head" area always have tailroom for skb_shared_info */ 5153 frame_sz = (void *)skb_end_pointer(skb) - hard_start; 5154 frame_sz += SKB_DATA_ALIGN(sizeof(struct skb_shared_info)); 5155 5156 rxqueue = netif_get_rxqueue(skb); 5157 xdp_init_buff(xdp, frame_sz, &rxqueue->xdp_rxq); 5158 xdp_prepare_buff(xdp, hard_start, skb_headroom(skb) - mac_len, 5159 skb_headlen(skb) + mac_len, true); 5160 if (skb_is_nonlinear(skb)) { 5161 skb_shinfo(skb)->xdp_frags_size = skb->data_len; 5162 xdp_buff_set_frags_flag(xdp); 5163 } else { 5164 xdp_buff_clear_frags_flag(xdp); 5165 } 5166 5167 orig_data_end = xdp->data_end; 5168 orig_data = xdp->data; 5169 eth = (struct ethhdr *)xdp->data; 5170 orig_host = ether_addr_equal_64bits(eth->h_dest, skb->dev->dev_addr); 5171 orig_bcast = is_multicast_ether_addr_64bits(eth->h_dest); 5172 orig_eth_type = eth->h_proto; 5173 5174 act = bpf_prog_run_xdp(xdp_prog, xdp); 5175 5176 /* check if bpf_xdp_adjust_head was used */ 5177 off = xdp->data - orig_data; 5178 if (off) { 5179 if (off > 0) 5180 __skb_pull(skb, off); 5181 else if (off < 0) 5182 __skb_push(skb, -off); 5183 5184 skb->mac_header += off; 5185 skb_reset_network_header(skb); 5186 } 5187 5188 /* check if bpf_xdp_adjust_tail was used */ 5189 off = xdp->data_end - orig_data_end; 5190 if (off != 0) { 5191 skb_set_tail_pointer(skb, xdp->data_end - xdp->data); 5192 skb->len += off; /* positive on grow, negative on shrink */ 5193 } 5194 5195 /* XDP frag metadata (e.g. nr_frags) are updated in eBPF helpers 5196 * (e.g. bpf_xdp_adjust_tail), we need to update data_len here. 5197 */ 5198 if (xdp_buff_has_frags(xdp)) 5199 skb->data_len = skb_shinfo(skb)->xdp_frags_size; 5200 else 5201 skb->data_len = 0; 5202 5203 /* check if XDP changed eth hdr such SKB needs update */ 5204 eth = (struct ethhdr *)xdp->data; 5205 if ((orig_eth_type != eth->h_proto) || 5206 (orig_host != ether_addr_equal_64bits(eth->h_dest, 5207 skb->dev->dev_addr)) || 5208 (orig_bcast != is_multicast_ether_addr_64bits(eth->h_dest))) { 5209 __skb_push(skb, ETH_HLEN); 5210 skb->pkt_type = PACKET_HOST; 5211 skb->protocol = eth_type_trans(skb, skb->dev); 5212 } 5213 5214 /* Redirect/Tx gives L2 packet, code that will reuse skb must __skb_pull 5215 * before calling us again on redirect path. We do not call do_redirect 5216 * as we leave that up to the caller. 5217 * 5218 * Caller is responsible for managing lifetime of skb (i.e. calling 5219 * kfree_skb in response to actions it cannot handle/XDP_DROP). 5220 */ 5221 switch (act) { 5222 case XDP_REDIRECT: 5223 case XDP_TX: 5224 __skb_push(skb, mac_len); 5225 break; 5226 case XDP_PASS: 5227 metalen = xdp->data - xdp->data_meta; 5228 if (metalen) 5229 skb_metadata_set(skb, metalen); 5230 break; 5231 } 5232 5233 return act; 5234 } 5235 5236 static int 5237 netif_skb_check_for_xdp(struct sk_buff **pskb, const struct bpf_prog *prog) 5238 { 5239 struct sk_buff *skb = *pskb; 5240 int err, hroom, troom; 5241 5242 if (!skb_cow_data_for_xdp(this_cpu_read(system_page_pool), pskb, prog)) 5243 return 0; 5244 5245 /* In case we have to go down the path and also linearize, 5246 * then lets do the pskb_expand_head() work just once here. 5247 */ 5248 hroom = XDP_PACKET_HEADROOM - skb_headroom(skb); 5249 troom = skb->tail + skb->data_len - skb->end; 5250 err = pskb_expand_head(skb, 5251 hroom > 0 ? ALIGN(hroom, NET_SKB_PAD) : 0, 5252 troom > 0 ? troom + 128 : 0, GFP_ATOMIC); 5253 if (err) 5254 return err; 5255 5256 return skb_linearize(skb); 5257 } 5258 5259 static u32 netif_receive_generic_xdp(struct sk_buff **pskb, 5260 struct xdp_buff *xdp, 5261 const struct bpf_prog *xdp_prog) 5262 { 5263 struct sk_buff *skb = *pskb; 5264 u32 mac_len, act = XDP_DROP; 5265 5266 /* Reinjected packets coming from act_mirred or similar should 5267 * not get XDP generic processing. 5268 */ 5269 if (skb_is_redirected(skb)) 5270 return XDP_PASS; 5271 5272 /* XDP packets must have sufficient headroom of XDP_PACKET_HEADROOM 5273 * bytes. This is the guarantee that also native XDP provides, 5274 * thus we need to do it here as well. 5275 */ 5276 mac_len = skb->data - skb_mac_header(skb); 5277 __skb_push(skb, mac_len); 5278 5279 if (skb_cloned(skb) || skb_is_nonlinear(skb) || 5280 skb_headroom(skb) < XDP_PACKET_HEADROOM) { 5281 if (netif_skb_check_for_xdp(pskb, xdp_prog)) 5282 goto do_drop; 5283 } 5284 5285 __skb_pull(*pskb, mac_len); 5286 5287 act = bpf_prog_run_generic_xdp(*pskb, xdp, xdp_prog); 5288 switch (act) { 5289 case XDP_REDIRECT: 5290 case XDP_TX: 5291 case XDP_PASS: 5292 break; 5293 default: 5294 bpf_warn_invalid_xdp_action((*pskb)->dev, xdp_prog, act); 5295 fallthrough; 5296 case XDP_ABORTED: 5297 trace_xdp_exception((*pskb)->dev, xdp_prog, act); 5298 fallthrough; 5299 case XDP_DROP: 5300 do_drop: 5301 kfree_skb(*pskb); 5302 break; 5303 } 5304 5305 return act; 5306 } 5307 5308 /* When doing generic XDP we have to bypass the qdisc layer and the 5309 * network taps in order to match in-driver-XDP behavior. This also means 5310 * that XDP packets are able to starve other packets going through a qdisc, 5311 * and DDOS attacks will be more effective. In-driver-XDP use dedicated TX 5312 * queues, so they do not have this starvation issue. 5313 */ 5314 void generic_xdp_tx(struct sk_buff *skb, const struct bpf_prog *xdp_prog) 5315 { 5316 struct net_device *dev = skb->dev; 5317 struct netdev_queue *txq; 5318 bool free_skb = true; 5319 int cpu, rc; 5320 5321 txq = netdev_core_pick_tx(dev, skb, NULL); 5322 cpu = smp_processor_id(); 5323 HARD_TX_LOCK(dev, txq, cpu); 5324 if (!netif_xmit_frozen_or_drv_stopped(txq)) { 5325 rc = netdev_start_xmit(skb, dev, txq, 0); 5326 if (dev_xmit_complete(rc)) 5327 free_skb = false; 5328 } 5329 HARD_TX_UNLOCK(dev, txq); 5330 if (free_skb) { 5331 trace_xdp_exception(dev, xdp_prog, XDP_TX); 5332 dev_core_stats_tx_dropped_inc(dev); 5333 kfree_skb(skb); 5334 } 5335 } 5336 5337 static DEFINE_STATIC_KEY_FALSE(generic_xdp_needed_key); 5338 5339 int do_xdp_generic(const struct bpf_prog *xdp_prog, struct sk_buff **pskb) 5340 { 5341 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 5342 5343 if (xdp_prog) { 5344 struct xdp_buff xdp; 5345 u32 act; 5346 int err; 5347 5348 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 5349 act = netif_receive_generic_xdp(pskb, &xdp, xdp_prog); 5350 if (act != XDP_PASS) { 5351 switch (act) { 5352 case XDP_REDIRECT: 5353 err = xdp_do_generic_redirect((*pskb)->dev, *pskb, 5354 &xdp, xdp_prog); 5355 if (err) 5356 goto out_redir; 5357 break; 5358 case XDP_TX: 5359 generic_xdp_tx(*pskb, xdp_prog); 5360 break; 5361 } 5362 bpf_net_ctx_clear(bpf_net_ctx); 5363 return XDP_DROP; 5364 } 5365 bpf_net_ctx_clear(bpf_net_ctx); 5366 } 5367 return XDP_PASS; 5368 out_redir: 5369 bpf_net_ctx_clear(bpf_net_ctx); 5370 kfree_skb_reason(*pskb, SKB_DROP_REASON_XDP); 5371 return XDP_DROP; 5372 } 5373 EXPORT_SYMBOL_GPL(do_xdp_generic); 5374 5375 static int netif_rx_internal(struct sk_buff *skb) 5376 { 5377 int ret; 5378 5379 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), skb); 5380 5381 trace_netif_rx(skb); 5382 5383 #ifdef CONFIG_RPS 5384 if (static_branch_unlikely(&rps_needed)) { 5385 struct rps_dev_flow voidflow, *rflow = &voidflow; 5386 int cpu; 5387 5388 rcu_read_lock(); 5389 5390 cpu = get_rps_cpu(skb->dev, skb, &rflow); 5391 if (cpu < 0) 5392 cpu = smp_processor_id(); 5393 5394 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 5395 5396 rcu_read_unlock(); 5397 } else 5398 #endif 5399 { 5400 unsigned int qtail; 5401 5402 ret = enqueue_to_backlog(skb, smp_processor_id(), &qtail); 5403 } 5404 return ret; 5405 } 5406 5407 /** 5408 * __netif_rx - Slightly optimized version of netif_rx 5409 * @skb: buffer to post 5410 * 5411 * This behaves as netif_rx except that it does not disable bottom halves. 5412 * As a result this function may only be invoked from the interrupt context 5413 * (either hard or soft interrupt). 5414 */ 5415 int __netif_rx(struct sk_buff *skb) 5416 { 5417 int ret; 5418 5419 lockdep_assert_once(hardirq_count() | softirq_count()); 5420 5421 trace_netif_rx_entry(skb); 5422 ret = netif_rx_internal(skb); 5423 trace_netif_rx_exit(ret); 5424 return ret; 5425 } 5426 EXPORT_SYMBOL(__netif_rx); 5427 5428 /** 5429 * netif_rx - post buffer to the network code 5430 * @skb: buffer to post 5431 * 5432 * This function receives a packet from a device driver and queues it for 5433 * the upper (protocol) levels to process via the backlog NAPI device. It 5434 * always succeeds. The buffer may be dropped during processing for 5435 * congestion control or by the protocol layers. 5436 * The network buffer is passed via the backlog NAPI device. Modern NIC 5437 * driver should use NAPI and GRO. 5438 * This function can used from interrupt and from process context. The 5439 * caller from process context must not disable interrupts before invoking 5440 * this function. 5441 * 5442 * return values: 5443 * NET_RX_SUCCESS (no congestion) 5444 * NET_RX_DROP (packet was dropped) 5445 * 5446 */ 5447 int netif_rx(struct sk_buff *skb) 5448 { 5449 bool need_bh_off = !(hardirq_count() | softirq_count()); 5450 int ret; 5451 5452 if (need_bh_off) 5453 local_bh_disable(); 5454 trace_netif_rx_entry(skb); 5455 ret = netif_rx_internal(skb); 5456 trace_netif_rx_exit(ret); 5457 if (need_bh_off) 5458 local_bh_enable(); 5459 return ret; 5460 } 5461 EXPORT_SYMBOL(netif_rx); 5462 5463 static __latent_entropy void net_tx_action(void) 5464 { 5465 struct softnet_data *sd = this_cpu_ptr(&softnet_data); 5466 5467 if (sd->completion_queue) { 5468 struct sk_buff *clist; 5469 5470 local_irq_disable(); 5471 clist = sd->completion_queue; 5472 sd->completion_queue = NULL; 5473 local_irq_enable(); 5474 5475 while (clist) { 5476 struct sk_buff *skb = clist; 5477 5478 clist = clist->next; 5479 5480 WARN_ON(refcount_read(&skb->users)); 5481 if (likely(get_kfree_skb_cb(skb)->reason == SKB_CONSUMED)) 5482 trace_consume_skb(skb, net_tx_action); 5483 else 5484 trace_kfree_skb(skb, net_tx_action, 5485 get_kfree_skb_cb(skb)->reason, NULL); 5486 5487 if (skb->fclone != SKB_FCLONE_UNAVAILABLE) 5488 __kfree_skb(skb); 5489 else 5490 __napi_kfree_skb(skb, 5491 get_kfree_skb_cb(skb)->reason); 5492 } 5493 } 5494 5495 if (sd->output_queue) { 5496 struct Qdisc *head; 5497 5498 local_irq_disable(); 5499 head = sd->output_queue; 5500 sd->output_queue = NULL; 5501 sd->output_queue_tailp = &sd->output_queue; 5502 local_irq_enable(); 5503 5504 rcu_read_lock(); 5505 5506 while (head) { 5507 struct Qdisc *q = head; 5508 spinlock_t *root_lock = NULL; 5509 5510 head = head->next_sched; 5511 5512 /* We need to make sure head->next_sched is read 5513 * before clearing __QDISC_STATE_SCHED 5514 */ 5515 smp_mb__before_atomic(); 5516 5517 if (!(q->flags & TCQ_F_NOLOCK)) { 5518 root_lock = qdisc_lock(q); 5519 spin_lock(root_lock); 5520 } else if (unlikely(test_bit(__QDISC_STATE_DEACTIVATED, 5521 &q->state))) { 5522 /* There is a synchronize_net() between 5523 * STATE_DEACTIVATED flag being set and 5524 * qdisc_reset()/some_qdisc_is_busy() in 5525 * dev_deactivate(), so we can safely bail out 5526 * early here to avoid data race between 5527 * qdisc_deactivate() and some_qdisc_is_busy() 5528 * for lockless qdisc. 5529 */ 5530 clear_bit(__QDISC_STATE_SCHED, &q->state); 5531 continue; 5532 } 5533 5534 clear_bit(__QDISC_STATE_SCHED, &q->state); 5535 qdisc_run(q); 5536 if (root_lock) 5537 spin_unlock(root_lock); 5538 } 5539 5540 rcu_read_unlock(); 5541 } 5542 5543 xfrm_dev_backlog(sd); 5544 } 5545 5546 #if IS_ENABLED(CONFIG_BRIDGE) && IS_ENABLED(CONFIG_ATM_LANE) 5547 /* This hook is defined here for ATM LANE */ 5548 int (*br_fdb_test_addr_hook)(struct net_device *dev, 5549 unsigned char *addr) __read_mostly; 5550 EXPORT_SYMBOL_GPL(br_fdb_test_addr_hook); 5551 #endif 5552 5553 /** 5554 * netdev_is_rx_handler_busy - check if receive handler is registered 5555 * @dev: device to check 5556 * 5557 * Check if a receive handler is already registered for a given device. 5558 * Return true if there one. 5559 * 5560 * The caller must hold the rtnl_mutex. 5561 */ 5562 bool netdev_is_rx_handler_busy(struct net_device *dev) 5563 { 5564 ASSERT_RTNL(); 5565 return dev && rtnl_dereference(dev->rx_handler); 5566 } 5567 EXPORT_SYMBOL_GPL(netdev_is_rx_handler_busy); 5568 5569 /** 5570 * netdev_rx_handler_register - register receive handler 5571 * @dev: device to register a handler for 5572 * @rx_handler: receive handler to register 5573 * @rx_handler_data: data pointer that is used by rx handler 5574 * 5575 * Register a receive handler for a device. This handler will then be 5576 * called from __netif_receive_skb. A negative errno code is returned 5577 * on a failure. 5578 * 5579 * The caller must hold the rtnl_mutex. 5580 * 5581 * For a general description of rx_handler, see enum rx_handler_result. 5582 */ 5583 int netdev_rx_handler_register(struct net_device *dev, 5584 rx_handler_func_t *rx_handler, 5585 void *rx_handler_data) 5586 { 5587 if (netdev_is_rx_handler_busy(dev)) 5588 return -EBUSY; 5589 5590 if (dev->priv_flags & IFF_NO_RX_HANDLER) 5591 return -EINVAL; 5592 5593 /* Note: rx_handler_data must be set before rx_handler */ 5594 rcu_assign_pointer(dev->rx_handler_data, rx_handler_data); 5595 rcu_assign_pointer(dev->rx_handler, rx_handler); 5596 5597 return 0; 5598 } 5599 EXPORT_SYMBOL_GPL(netdev_rx_handler_register); 5600 5601 /** 5602 * netdev_rx_handler_unregister - unregister receive handler 5603 * @dev: device to unregister a handler from 5604 * 5605 * Unregister a receive handler from a device. 5606 * 5607 * The caller must hold the rtnl_mutex. 5608 */ 5609 void netdev_rx_handler_unregister(struct net_device *dev) 5610 { 5611 5612 ASSERT_RTNL(); 5613 RCU_INIT_POINTER(dev->rx_handler, NULL); 5614 /* a reader seeing a non NULL rx_handler in a rcu_read_lock() 5615 * section has a guarantee to see a non NULL rx_handler_data 5616 * as well. 5617 */ 5618 synchronize_net(); 5619 RCU_INIT_POINTER(dev->rx_handler_data, NULL); 5620 } 5621 EXPORT_SYMBOL_GPL(netdev_rx_handler_unregister); 5622 5623 /* 5624 * Limit the use of PFMEMALLOC reserves to those protocols that implement 5625 * the special handling of PFMEMALLOC skbs. 5626 */ 5627 static bool skb_pfmemalloc_protocol(struct sk_buff *skb) 5628 { 5629 switch (skb->protocol) { 5630 case htons(ETH_P_ARP): 5631 case htons(ETH_P_IP): 5632 case htons(ETH_P_IPV6): 5633 case htons(ETH_P_8021Q): 5634 case htons(ETH_P_8021AD): 5635 return true; 5636 default: 5637 return false; 5638 } 5639 } 5640 5641 static inline int nf_ingress(struct sk_buff *skb, struct packet_type **pt_prev, 5642 int *ret, struct net_device *orig_dev) 5643 { 5644 if (nf_hook_ingress_active(skb)) { 5645 int ingress_retval; 5646 5647 if (*pt_prev) { 5648 *ret = deliver_skb(skb, *pt_prev, orig_dev); 5649 *pt_prev = NULL; 5650 } 5651 5652 rcu_read_lock(); 5653 ingress_retval = nf_hook_ingress(skb); 5654 rcu_read_unlock(); 5655 return ingress_retval; 5656 } 5657 return 0; 5658 } 5659 5660 static int __netif_receive_skb_core(struct sk_buff **pskb, bool pfmemalloc, 5661 struct packet_type **ppt_prev) 5662 { 5663 struct packet_type *ptype, *pt_prev; 5664 rx_handler_func_t *rx_handler; 5665 struct sk_buff *skb = *pskb; 5666 struct net_device *orig_dev; 5667 bool deliver_exact = false; 5668 int ret = NET_RX_DROP; 5669 __be16 type; 5670 5671 net_timestamp_check(!READ_ONCE(net_hotdata.tstamp_prequeue), skb); 5672 5673 trace_netif_receive_skb(skb); 5674 5675 orig_dev = skb->dev; 5676 5677 skb_reset_network_header(skb); 5678 #if !defined(CONFIG_DEBUG_NET) 5679 /* We plan to no longer reset the transport header here. 5680 * Give some time to fuzzers and dev build to catch bugs 5681 * in network stacks. 5682 */ 5683 if (!skb_transport_header_was_set(skb)) 5684 skb_reset_transport_header(skb); 5685 #endif 5686 skb_reset_mac_len(skb); 5687 5688 pt_prev = NULL; 5689 5690 another_round: 5691 skb->skb_iif = skb->dev->ifindex; 5692 5693 __this_cpu_inc(softnet_data.processed); 5694 5695 if (static_branch_unlikely(&generic_xdp_needed_key)) { 5696 int ret2; 5697 5698 migrate_disable(); 5699 ret2 = do_xdp_generic(rcu_dereference(skb->dev->xdp_prog), 5700 &skb); 5701 migrate_enable(); 5702 5703 if (ret2 != XDP_PASS) { 5704 ret = NET_RX_DROP; 5705 goto out; 5706 } 5707 } 5708 5709 if (eth_type_vlan(skb->protocol)) { 5710 skb = skb_vlan_untag(skb); 5711 if (unlikely(!skb)) 5712 goto out; 5713 } 5714 5715 if (skb_skip_tc_classify(skb)) 5716 goto skip_classify; 5717 5718 if (pfmemalloc) 5719 goto skip_taps; 5720 5721 list_for_each_entry_rcu(ptype, &net_hotdata.ptype_all, list) { 5722 if (pt_prev) 5723 ret = deliver_skb(skb, pt_prev, orig_dev); 5724 pt_prev = ptype; 5725 } 5726 5727 list_for_each_entry_rcu(ptype, &skb->dev->ptype_all, list) { 5728 if (pt_prev) 5729 ret = deliver_skb(skb, pt_prev, orig_dev); 5730 pt_prev = ptype; 5731 } 5732 5733 skip_taps: 5734 #ifdef CONFIG_NET_INGRESS 5735 if (static_branch_unlikely(&ingress_needed_key)) { 5736 bool another = false; 5737 5738 nf_skip_egress(skb, true); 5739 skb = sch_handle_ingress(skb, &pt_prev, &ret, orig_dev, 5740 &another); 5741 if (another) 5742 goto another_round; 5743 if (!skb) 5744 goto out; 5745 5746 nf_skip_egress(skb, false); 5747 if (nf_ingress(skb, &pt_prev, &ret, orig_dev) < 0) 5748 goto out; 5749 } 5750 #endif 5751 skb_reset_redirect(skb); 5752 skip_classify: 5753 if (pfmemalloc && !skb_pfmemalloc_protocol(skb)) 5754 goto drop; 5755 5756 if (skb_vlan_tag_present(skb)) { 5757 if (pt_prev) { 5758 ret = deliver_skb(skb, pt_prev, orig_dev); 5759 pt_prev = NULL; 5760 } 5761 if (vlan_do_receive(&skb)) 5762 goto another_round; 5763 else if (unlikely(!skb)) 5764 goto out; 5765 } 5766 5767 rx_handler = rcu_dereference(skb->dev->rx_handler); 5768 if (rx_handler) { 5769 if (pt_prev) { 5770 ret = deliver_skb(skb, pt_prev, orig_dev); 5771 pt_prev = NULL; 5772 } 5773 switch (rx_handler(&skb)) { 5774 case RX_HANDLER_CONSUMED: 5775 ret = NET_RX_SUCCESS; 5776 goto out; 5777 case RX_HANDLER_ANOTHER: 5778 goto another_round; 5779 case RX_HANDLER_EXACT: 5780 deliver_exact = true; 5781 break; 5782 case RX_HANDLER_PASS: 5783 break; 5784 default: 5785 BUG(); 5786 } 5787 } 5788 5789 if (unlikely(skb_vlan_tag_present(skb)) && !netdev_uses_dsa(skb->dev)) { 5790 check_vlan_id: 5791 if (skb_vlan_tag_get_id(skb)) { 5792 /* Vlan id is non 0 and vlan_do_receive() above couldn't 5793 * find vlan device. 5794 */ 5795 skb->pkt_type = PACKET_OTHERHOST; 5796 } else if (eth_type_vlan(skb->protocol)) { 5797 /* Outer header is 802.1P with vlan 0, inner header is 5798 * 802.1Q or 802.1AD and vlan_do_receive() above could 5799 * not find vlan dev for vlan id 0. 5800 */ 5801 __vlan_hwaccel_clear_tag(skb); 5802 skb = skb_vlan_untag(skb); 5803 if (unlikely(!skb)) 5804 goto out; 5805 if (vlan_do_receive(&skb)) 5806 /* After stripping off 802.1P header with vlan 0 5807 * vlan dev is found for inner header. 5808 */ 5809 goto another_round; 5810 else if (unlikely(!skb)) 5811 goto out; 5812 else 5813 /* We have stripped outer 802.1P vlan 0 header. 5814 * But could not find vlan dev. 5815 * check again for vlan id to set OTHERHOST. 5816 */ 5817 goto check_vlan_id; 5818 } 5819 /* Note: we might in the future use prio bits 5820 * and set skb->priority like in vlan_do_receive() 5821 * For the time being, just ignore Priority Code Point 5822 */ 5823 __vlan_hwaccel_clear_tag(skb); 5824 } 5825 5826 type = skb->protocol; 5827 5828 /* deliver only exact match when indicated */ 5829 if (likely(!deliver_exact)) { 5830 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 5831 &ptype_base[ntohs(type) & 5832 PTYPE_HASH_MASK]); 5833 } 5834 5835 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 5836 &orig_dev->ptype_specific); 5837 5838 if (unlikely(skb->dev != orig_dev)) { 5839 deliver_ptype_list_skb(skb, &pt_prev, orig_dev, type, 5840 &skb->dev->ptype_specific); 5841 } 5842 5843 if (pt_prev) { 5844 if (unlikely(skb_orphan_frags_rx(skb, GFP_ATOMIC))) 5845 goto drop; 5846 *ppt_prev = pt_prev; 5847 } else { 5848 drop: 5849 if (!deliver_exact) 5850 dev_core_stats_rx_dropped_inc(skb->dev); 5851 else 5852 dev_core_stats_rx_nohandler_inc(skb->dev); 5853 kfree_skb_reason(skb, SKB_DROP_REASON_UNHANDLED_PROTO); 5854 /* Jamal, now you will not able to escape explaining 5855 * me how you were going to use this. :-) 5856 */ 5857 ret = NET_RX_DROP; 5858 } 5859 5860 out: 5861 /* The invariant here is that if *ppt_prev is not NULL 5862 * then skb should also be non-NULL. 5863 * 5864 * Apparently *ppt_prev assignment above holds this invariant due to 5865 * skb dereferencing near it. 5866 */ 5867 *pskb = skb; 5868 return ret; 5869 } 5870 5871 static int __netif_receive_skb_one_core(struct sk_buff *skb, bool pfmemalloc) 5872 { 5873 struct net_device *orig_dev = skb->dev; 5874 struct packet_type *pt_prev = NULL; 5875 int ret; 5876 5877 ret = __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev); 5878 if (pt_prev) 5879 ret = INDIRECT_CALL_INET(pt_prev->func, ipv6_rcv, ip_rcv, skb, 5880 skb->dev, pt_prev, orig_dev); 5881 return ret; 5882 } 5883 5884 /** 5885 * netif_receive_skb_core - special purpose version of netif_receive_skb 5886 * @skb: buffer to process 5887 * 5888 * More direct receive version of netif_receive_skb(). It should 5889 * only be used by callers that have a need to skip RPS and Generic XDP. 5890 * Caller must also take care of handling if ``(page_is_)pfmemalloc``. 5891 * 5892 * This function may only be called from softirq context and interrupts 5893 * should be enabled. 5894 * 5895 * Return values (usually ignored): 5896 * NET_RX_SUCCESS: no congestion 5897 * NET_RX_DROP: packet was dropped 5898 */ 5899 int netif_receive_skb_core(struct sk_buff *skb) 5900 { 5901 int ret; 5902 5903 rcu_read_lock(); 5904 ret = __netif_receive_skb_one_core(skb, false); 5905 rcu_read_unlock(); 5906 5907 return ret; 5908 } 5909 EXPORT_SYMBOL(netif_receive_skb_core); 5910 5911 static inline void __netif_receive_skb_list_ptype(struct list_head *head, 5912 struct packet_type *pt_prev, 5913 struct net_device *orig_dev) 5914 { 5915 struct sk_buff *skb, *next; 5916 5917 if (!pt_prev) 5918 return; 5919 if (list_empty(head)) 5920 return; 5921 if (pt_prev->list_func != NULL) 5922 INDIRECT_CALL_INET(pt_prev->list_func, ipv6_list_rcv, 5923 ip_list_rcv, head, pt_prev, orig_dev); 5924 else 5925 list_for_each_entry_safe(skb, next, head, list) { 5926 skb_list_del_init(skb); 5927 pt_prev->func(skb, skb->dev, pt_prev, orig_dev); 5928 } 5929 } 5930 5931 static void __netif_receive_skb_list_core(struct list_head *head, bool pfmemalloc) 5932 { 5933 /* Fast-path assumptions: 5934 * - There is no RX handler. 5935 * - Only one packet_type matches. 5936 * If either of these fails, we will end up doing some per-packet 5937 * processing in-line, then handling the 'last ptype' for the whole 5938 * sublist. This can't cause out-of-order delivery to any single ptype, 5939 * because the 'last ptype' must be constant across the sublist, and all 5940 * other ptypes are handled per-packet. 5941 */ 5942 /* Current (common) ptype of sublist */ 5943 struct packet_type *pt_curr = NULL; 5944 /* Current (common) orig_dev of sublist */ 5945 struct net_device *od_curr = NULL; 5946 struct sk_buff *skb, *next; 5947 LIST_HEAD(sublist); 5948 5949 list_for_each_entry_safe(skb, next, head, list) { 5950 struct net_device *orig_dev = skb->dev; 5951 struct packet_type *pt_prev = NULL; 5952 5953 skb_list_del_init(skb); 5954 __netif_receive_skb_core(&skb, pfmemalloc, &pt_prev); 5955 if (!pt_prev) 5956 continue; 5957 if (pt_curr != pt_prev || od_curr != orig_dev) { 5958 /* dispatch old sublist */ 5959 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr); 5960 /* start new sublist */ 5961 INIT_LIST_HEAD(&sublist); 5962 pt_curr = pt_prev; 5963 od_curr = orig_dev; 5964 } 5965 list_add_tail(&skb->list, &sublist); 5966 } 5967 5968 /* dispatch final sublist */ 5969 __netif_receive_skb_list_ptype(&sublist, pt_curr, od_curr); 5970 } 5971 5972 static int __netif_receive_skb(struct sk_buff *skb) 5973 { 5974 int ret; 5975 5976 if (sk_memalloc_socks() && skb_pfmemalloc(skb)) { 5977 unsigned int noreclaim_flag; 5978 5979 /* 5980 * PFMEMALLOC skbs are special, they should 5981 * - be delivered to SOCK_MEMALLOC sockets only 5982 * - stay away from userspace 5983 * - have bounded memory usage 5984 * 5985 * Use PF_MEMALLOC as this saves us from propagating the allocation 5986 * context down to all allocation sites. 5987 */ 5988 noreclaim_flag = memalloc_noreclaim_save(); 5989 ret = __netif_receive_skb_one_core(skb, true); 5990 memalloc_noreclaim_restore(noreclaim_flag); 5991 } else 5992 ret = __netif_receive_skb_one_core(skb, false); 5993 5994 return ret; 5995 } 5996 5997 static void __netif_receive_skb_list(struct list_head *head) 5998 { 5999 unsigned long noreclaim_flag = 0; 6000 struct sk_buff *skb, *next; 6001 bool pfmemalloc = false; /* Is current sublist PF_MEMALLOC? */ 6002 6003 list_for_each_entry_safe(skb, next, head, list) { 6004 if ((sk_memalloc_socks() && skb_pfmemalloc(skb)) != pfmemalloc) { 6005 struct list_head sublist; 6006 6007 /* Handle the previous sublist */ 6008 list_cut_before(&sublist, head, &skb->list); 6009 if (!list_empty(&sublist)) 6010 __netif_receive_skb_list_core(&sublist, pfmemalloc); 6011 pfmemalloc = !pfmemalloc; 6012 /* See comments in __netif_receive_skb */ 6013 if (pfmemalloc) 6014 noreclaim_flag = memalloc_noreclaim_save(); 6015 else 6016 memalloc_noreclaim_restore(noreclaim_flag); 6017 } 6018 } 6019 /* Handle the remaining sublist */ 6020 if (!list_empty(head)) 6021 __netif_receive_skb_list_core(head, pfmemalloc); 6022 /* Restore pflags */ 6023 if (pfmemalloc) 6024 memalloc_noreclaim_restore(noreclaim_flag); 6025 } 6026 6027 static int generic_xdp_install(struct net_device *dev, struct netdev_bpf *xdp) 6028 { 6029 struct bpf_prog *old = rtnl_dereference(dev->xdp_prog); 6030 struct bpf_prog *new = xdp->prog; 6031 int ret = 0; 6032 6033 switch (xdp->command) { 6034 case XDP_SETUP_PROG: 6035 rcu_assign_pointer(dev->xdp_prog, new); 6036 if (old) 6037 bpf_prog_put(old); 6038 6039 if (old && !new) { 6040 static_branch_dec(&generic_xdp_needed_key); 6041 } else if (new && !old) { 6042 static_branch_inc(&generic_xdp_needed_key); 6043 netif_disable_lro(dev); 6044 dev_disable_gro_hw(dev); 6045 } 6046 break; 6047 6048 default: 6049 ret = -EINVAL; 6050 break; 6051 } 6052 6053 return ret; 6054 } 6055 6056 static int netif_receive_skb_internal(struct sk_buff *skb) 6057 { 6058 int ret; 6059 6060 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), skb); 6061 6062 if (skb_defer_rx_timestamp(skb)) 6063 return NET_RX_SUCCESS; 6064 6065 rcu_read_lock(); 6066 #ifdef CONFIG_RPS 6067 if (static_branch_unlikely(&rps_needed)) { 6068 struct rps_dev_flow voidflow, *rflow = &voidflow; 6069 int cpu = get_rps_cpu(skb->dev, skb, &rflow); 6070 6071 if (cpu >= 0) { 6072 ret = enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 6073 rcu_read_unlock(); 6074 return ret; 6075 } 6076 } 6077 #endif 6078 ret = __netif_receive_skb(skb); 6079 rcu_read_unlock(); 6080 return ret; 6081 } 6082 6083 void netif_receive_skb_list_internal(struct list_head *head) 6084 { 6085 struct sk_buff *skb, *next; 6086 LIST_HEAD(sublist); 6087 6088 list_for_each_entry_safe(skb, next, head, list) { 6089 net_timestamp_check(READ_ONCE(net_hotdata.tstamp_prequeue), 6090 skb); 6091 skb_list_del_init(skb); 6092 if (!skb_defer_rx_timestamp(skb)) 6093 list_add_tail(&skb->list, &sublist); 6094 } 6095 list_splice_init(&sublist, head); 6096 6097 rcu_read_lock(); 6098 #ifdef CONFIG_RPS 6099 if (static_branch_unlikely(&rps_needed)) { 6100 list_for_each_entry_safe(skb, next, head, list) { 6101 struct rps_dev_flow voidflow, *rflow = &voidflow; 6102 int cpu = get_rps_cpu(skb->dev, skb, &rflow); 6103 6104 if (cpu >= 0) { 6105 /* Will be handled, remove from list */ 6106 skb_list_del_init(skb); 6107 enqueue_to_backlog(skb, cpu, &rflow->last_qtail); 6108 } 6109 } 6110 } 6111 #endif 6112 __netif_receive_skb_list(head); 6113 rcu_read_unlock(); 6114 } 6115 6116 /** 6117 * netif_receive_skb - process receive buffer from network 6118 * @skb: buffer to process 6119 * 6120 * netif_receive_skb() is the main receive data processing function. 6121 * It always succeeds. The buffer may be dropped during processing 6122 * for congestion control or by the protocol layers. 6123 * 6124 * This function may only be called from softirq context and interrupts 6125 * should be enabled. 6126 * 6127 * Return values (usually ignored): 6128 * NET_RX_SUCCESS: no congestion 6129 * NET_RX_DROP: packet was dropped 6130 */ 6131 int netif_receive_skb(struct sk_buff *skb) 6132 { 6133 int ret; 6134 6135 trace_netif_receive_skb_entry(skb); 6136 6137 ret = netif_receive_skb_internal(skb); 6138 trace_netif_receive_skb_exit(ret); 6139 6140 return ret; 6141 } 6142 EXPORT_SYMBOL(netif_receive_skb); 6143 6144 /** 6145 * netif_receive_skb_list - process many receive buffers from network 6146 * @head: list of skbs to process. 6147 * 6148 * Since return value of netif_receive_skb() is normally ignored, and 6149 * wouldn't be meaningful for a list, this function returns void. 6150 * 6151 * This function may only be called from softirq context and interrupts 6152 * should be enabled. 6153 */ 6154 void netif_receive_skb_list(struct list_head *head) 6155 { 6156 struct sk_buff *skb; 6157 6158 if (list_empty(head)) 6159 return; 6160 if (trace_netif_receive_skb_list_entry_enabled()) { 6161 list_for_each_entry(skb, head, list) 6162 trace_netif_receive_skb_list_entry(skb); 6163 } 6164 netif_receive_skb_list_internal(head); 6165 trace_netif_receive_skb_list_exit(0); 6166 } 6167 EXPORT_SYMBOL(netif_receive_skb_list); 6168 6169 /* Network device is going away, flush any packets still pending */ 6170 static void flush_backlog(struct work_struct *work) 6171 { 6172 struct sk_buff *skb, *tmp; 6173 struct sk_buff_head list; 6174 struct softnet_data *sd; 6175 6176 __skb_queue_head_init(&list); 6177 local_bh_disable(); 6178 sd = this_cpu_ptr(&softnet_data); 6179 6180 backlog_lock_irq_disable(sd); 6181 skb_queue_walk_safe(&sd->input_pkt_queue, skb, tmp) { 6182 if (READ_ONCE(skb->dev->reg_state) == NETREG_UNREGISTERING) { 6183 __skb_unlink(skb, &sd->input_pkt_queue); 6184 __skb_queue_tail(&list, skb); 6185 rps_input_queue_head_incr(sd); 6186 } 6187 } 6188 backlog_unlock_irq_enable(sd); 6189 6190 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6191 skb_queue_walk_safe(&sd->process_queue, skb, tmp) { 6192 if (READ_ONCE(skb->dev->reg_state) == NETREG_UNREGISTERING) { 6193 __skb_unlink(skb, &sd->process_queue); 6194 __skb_queue_tail(&list, skb); 6195 rps_input_queue_head_incr(sd); 6196 } 6197 } 6198 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6199 local_bh_enable(); 6200 6201 __skb_queue_purge_reason(&list, SKB_DROP_REASON_DEV_READY); 6202 } 6203 6204 static bool flush_required(int cpu) 6205 { 6206 #if IS_ENABLED(CONFIG_RPS) 6207 struct softnet_data *sd = &per_cpu(softnet_data, cpu); 6208 bool do_flush; 6209 6210 backlog_lock_irq_disable(sd); 6211 6212 /* as insertion into process_queue happens with the rps lock held, 6213 * process_queue access may race only with dequeue 6214 */ 6215 do_flush = !skb_queue_empty(&sd->input_pkt_queue) || 6216 !skb_queue_empty_lockless(&sd->process_queue); 6217 backlog_unlock_irq_enable(sd); 6218 6219 return do_flush; 6220 #endif 6221 /* without RPS we can't safely check input_pkt_queue: during a 6222 * concurrent remote skb_queue_splice() we can detect as empty both 6223 * input_pkt_queue and process_queue even if the latter could end-up 6224 * containing a lot of packets. 6225 */ 6226 return true; 6227 } 6228 6229 struct flush_backlogs { 6230 cpumask_t flush_cpus; 6231 struct work_struct w[]; 6232 }; 6233 6234 static struct flush_backlogs *flush_backlogs_alloc(void) 6235 { 6236 return kmalloc(struct_size_t(struct flush_backlogs, w, nr_cpu_ids), 6237 GFP_KERNEL); 6238 } 6239 6240 static struct flush_backlogs *flush_backlogs_fallback; 6241 static DEFINE_MUTEX(flush_backlogs_mutex); 6242 6243 static void flush_all_backlogs(void) 6244 { 6245 struct flush_backlogs *ptr = flush_backlogs_alloc(); 6246 unsigned int cpu; 6247 6248 if (!ptr) { 6249 mutex_lock(&flush_backlogs_mutex); 6250 ptr = flush_backlogs_fallback; 6251 } 6252 cpumask_clear(&ptr->flush_cpus); 6253 6254 cpus_read_lock(); 6255 6256 for_each_online_cpu(cpu) { 6257 if (flush_required(cpu)) { 6258 INIT_WORK(&ptr->w[cpu], flush_backlog); 6259 queue_work_on(cpu, system_highpri_wq, &ptr->w[cpu]); 6260 __cpumask_set_cpu(cpu, &ptr->flush_cpus); 6261 } 6262 } 6263 6264 /* we can have in flight packet[s] on the cpus we are not flushing, 6265 * synchronize_net() in unregister_netdevice_many() will take care of 6266 * them. 6267 */ 6268 for_each_cpu(cpu, &ptr->flush_cpus) 6269 flush_work(&ptr->w[cpu]); 6270 6271 cpus_read_unlock(); 6272 6273 if (ptr != flush_backlogs_fallback) 6274 kfree(ptr); 6275 else 6276 mutex_unlock(&flush_backlogs_mutex); 6277 } 6278 6279 static void net_rps_send_ipi(struct softnet_data *remsd) 6280 { 6281 #ifdef CONFIG_RPS 6282 while (remsd) { 6283 struct softnet_data *next = remsd->rps_ipi_next; 6284 6285 if (cpu_online(remsd->cpu)) 6286 smp_call_function_single_async(remsd->cpu, &remsd->csd); 6287 remsd = next; 6288 } 6289 #endif 6290 } 6291 6292 /* 6293 * net_rps_action_and_irq_enable sends any pending IPI's for rps. 6294 * Note: called with local irq disabled, but exits with local irq enabled. 6295 */ 6296 static void net_rps_action_and_irq_enable(struct softnet_data *sd) 6297 { 6298 #ifdef CONFIG_RPS 6299 struct softnet_data *remsd = sd->rps_ipi_list; 6300 6301 if (!use_backlog_threads() && remsd) { 6302 sd->rps_ipi_list = NULL; 6303 6304 local_irq_enable(); 6305 6306 /* Send pending IPI's to kick RPS processing on remote cpus. */ 6307 net_rps_send_ipi(remsd); 6308 } else 6309 #endif 6310 local_irq_enable(); 6311 } 6312 6313 static bool sd_has_rps_ipi_waiting(struct softnet_data *sd) 6314 { 6315 #ifdef CONFIG_RPS 6316 return !use_backlog_threads() && sd->rps_ipi_list; 6317 #else 6318 return false; 6319 #endif 6320 } 6321 6322 static int process_backlog(struct napi_struct *napi, int quota) 6323 { 6324 struct softnet_data *sd = container_of(napi, struct softnet_data, backlog); 6325 bool again = true; 6326 int work = 0; 6327 6328 /* Check if we have pending ipi, its better to send them now, 6329 * not waiting net_rx_action() end. 6330 */ 6331 if (sd_has_rps_ipi_waiting(sd)) { 6332 local_irq_disable(); 6333 net_rps_action_and_irq_enable(sd); 6334 } 6335 6336 napi->weight = READ_ONCE(net_hotdata.dev_rx_weight); 6337 while (again) { 6338 struct sk_buff *skb; 6339 6340 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6341 while ((skb = __skb_dequeue(&sd->process_queue))) { 6342 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6343 rcu_read_lock(); 6344 __netif_receive_skb(skb); 6345 rcu_read_unlock(); 6346 if (++work >= quota) { 6347 rps_input_queue_head_add(sd, work); 6348 return work; 6349 } 6350 6351 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6352 } 6353 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6354 6355 backlog_lock_irq_disable(sd); 6356 if (skb_queue_empty(&sd->input_pkt_queue)) { 6357 /* 6358 * Inline a custom version of __napi_complete(). 6359 * only current cpu owns and manipulates this napi, 6360 * and NAPI_STATE_SCHED is the only possible flag set 6361 * on backlog. 6362 * We can use a plain write instead of clear_bit(), 6363 * and we dont need an smp_mb() memory barrier. 6364 */ 6365 napi->state &= NAPIF_STATE_THREADED; 6366 again = false; 6367 } else { 6368 local_lock_nested_bh(&softnet_data.process_queue_bh_lock); 6369 skb_queue_splice_tail_init(&sd->input_pkt_queue, 6370 &sd->process_queue); 6371 local_unlock_nested_bh(&softnet_data.process_queue_bh_lock); 6372 } 6373 backlog_unlock_irq_enable(sd); 6374 } 6375 6376 if (work) 6377 rps_input_queue_head_add(sd, work); 6378 return work; 6379 } 6380 6381 /** 6382 * __napi_schedule - schedule for receive 6383 * @n: entry to schedule 6384 * 6385 * The entry's receive function will be scheduled to run. 6386 * Consider using __napi_schedule_irqoff() if hard irqs are masked. 6387 */ 6388 void __napi_schedule(struct napi_struct *n) 6389 { 6390 unsigned long flags; 6391 6392 local_irq_save(flags); 6393 ____napi_schedule(this_cpu_ptr(&softnet_data), n); 6394 local_irq_restore(flags); 6395 } 6396 EXPORT_SYMBOL(__napi_schedule); 6397 6398 /** 6399 * napi_schedule_prep - check if napi can be scheduled 6400 * @n: napi context 6401 * 6402 * Test if NAPI routine is already running, and if not mark 6403 * it as running. This is used as a condition variable to 6404 * insure only one NAPI poll instance runs. We also make 6405 * sure there is no pending NAPI disable. 6406 */ 6407 bool napi_schedule_prep(struct napi_struct *n) 6408 { 6409 unsigned long new, val = READ_ONCE(n->state); 6410 6411 do { 6412 if (unlikely(val & NAPIF_STATE_DISABLE)) 6413 return false; 6414 new = val | NAPIF_STATE_SCHED; 6415 6416 /* Sets STATE_MISSED bit if STATE_SCHED was already set 6417 * This was suggested by Alexander Duyck, as compiler 6418 * emits better code than : 6419 * if (val & NAPIF_STATE_SCHED) 6420 * new |= NAPIF_STATE_MISSED; 6421 */ 6422 new |= (val & NAPIF_STATE_SCHED) / NAPIF_STATE_SCHED * 6423 NAPIF_STATE_MISSED; 6424 } while (!try_cmpxchg(&n->state, &val, new)); 6425 6426 return !(val & NAPIF_STATE_SCHED); 6427 } 6428 EXPORT_SYMBOL(napi_schedule_prep); 6429 6430 /** 6431 * __napi_schedule_irqoff - schedule for receive 6432 * @n: entry to schedule 6433 * 6434 * Variant of __napi_schedule() assuming hard irqs are masked. 6435 * 6436 * On PREEMPT_RT enabled kernels this maps to __napi_schedule() 6437 * because the interrupt disabled assumption might not be true 6438 * due to force-threaded interrupts and spinlock substitution. 6439 */ 6440 void __napi_schedule_irqoff(struct napi_struct *n) 6441 { 6442 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6443 ____napi_schedule(this_cpu_ptr(&softnet_data), n); 6444 else 6445 __napi_schedule(n); 6446 } 6447 EXPORT_SYMBOL(__napi_schedule_irqoff); 6448 6449 bool napi_complete_done(struct napi_struct *n, int work_done) 6450 { 6451 unsigned long flags, val, new, timeout = 0; 6452 bool ret = true; 6453 6454 /* 6455 * 1) Don't let napi dequeue from the cpu poll list 6456 * just in case its running on a different cpu. 6457 * 2) If we are busy polling, do nothing here, we have 6458 * the guarantee we will be called later. 6459 */ 6460 if (unlikely(n->state & (NAPIF_STATE_NPSVC | 6461 NAPIF_STATE_IN_BUSY_POLL))) 6462 return false; 6463 6464 if (work_done) { 6465 if (n->gro.bitmask) 6466 timeout = napi_get_gro_flush_timeout(n); 6467 n->defer_hard_irqs_count = napi_get_defer_hard_irqs(n); 6468 } 6469 if (n->defer_hard_irqs_count > 0) { 6470 n->defer_hard_irqs_count--; 6471 timeout = napi_get_gro_flush_timeout(n); 6472 if (timeout) 6473 ret = false; 6474 } 6475 6476 /* 6477 * When the NAPI instance uses a timeout and keeps postponing 6478 * it, we need to bound somehow the time packets are kept in 6479 * the GRO layer. 6480 */ 6481 gro_flush(&n->gro, !!timeout); 6482 gro_normal_list(&n->gro); 6483 6484 if (unlikely(!list_empty(&n->poll_list))) { 6485 /* If n->poll_list is not empty, we need to mask irqs */ 6486 local_irq_save(flags); 6487 list_del_init(&n->poll_list); 6488 local_irq_restore(flags); 6489 } 6490 WRITE_ONCE(n->list_owner, -1); 6491 6492 val = READ_ONCE(n->state); 6493 do { 6494 WARN_ON_ONCE(!(val & NAPIF_STATE_SCHED)); 6495 6496 new = val & ~(NAPIF_STATE_MISSED | NAPIF_STATE_SCHED | 6497 NAPIF_STATE_SCHED_THREADED | 6498 NAPIF_STATE_PREFER_BUSY_POLL); 6499 6500 /* If STATE_MISSED was set, leave STATE_SCHED set, 6501 * because we will call napi->poll() one more time. 6502 * This C code was suggested by Alexander Duyck to help gcc. 6503 */ 6504 new |= (val & NAPIF_STATE_MISSED) / NAPIF_STATE_MISSED * 6505 NAPIF_STATE_SCHED; 6506 } while (!try_cmpxchg(&n->state, &val, new)); 6507 6508 if (unlikely(val & NAPIF_STATE_MISSED)) { 6509 __napi_schedule(n); 6510 return false; 6511 } 6512 6513 if (timeout) 6514 hrtimer_start(&n->timer, ns_to_ktime(timeout), 6515 HRTIMER_MODE_REL_PINNED); 6516 return ret; 6517 } 6518 EXPORT_SYMBOL(napi_complete_done); 6519 6520 static void skb_defer_free_flush(struct softnet_data *sd) 6521 { 6522 struct sk_buff *skb, *next; 6523 6524 /* Paired with WRITE_ONCE() in skb_attempt_defer_free() */ 6525 if (!READ_ONCE(sd->defer_list)) 6526 return; 6527 6528 spin_lock(&sd->defer_lock); 6529 skb = sd->defer_list; 6530 sd->defer_list = NULL; 6531 sd->defer_count = 0; 6532 spin_unlock(&sd->defer_lock); 6533 6534 while (skb != NULL) { 6535 next = skb->next; 6536 napi_consume_skb(skb, 1); 6537 skb = next; 6538 } 6539 } 6540 6541 #if defined(CONFIG_NET_RX_BUSY_POLL) 6542 6543 static void __busy_poll_stop(struct napi_struct *napi, bool skip_schedule) 6544 { 6545 if (!skip_schedule) { 6546 gro_normal_list(&napi->gro); 6547 __napi_schedule(napi); 6548 return; 6549 } 6550 6551 /* Flush too old packets. If HZ < 1000, flush all packets */ 6552 gro_flush(&napi->gro, HZ >= 1000); 6553 gro_normal_list(&napi->gro); 6554 6555 clear_bit(NAPI_STATE_SCHED, &napi->state); 6556 } 6557 6558 enum { 6559 NAPI_F_PREFER_BUSY_POLL = 1, 6560 NAPI_F_END_ON_RESCHED = 2, 6561 }; 6562 6563 static void busy_poll_stop(struct napi_struct *napi, void *have_poll_lock, 6564 unsigned flags, u16 budget) 6565 { 6566 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 6567 bool skip_schedule = false; 6568 unsigned long timeout; 6569 int rc; 6570 6571 /* Busy polling means there is a high chance device driver hard irq 6572 * could not grab NAPI_STATE_SCHED, and that NAPI_STATE_MISSED was 6573 * set in napi_schedule_prep(). 6574 * Since we are about to call napi->poll() once more, we can safely 6575 * clear NAPI_STATE_MISSED. 6576 * 6577 * Note: x86 could use a single "lock and ..." instruction 6578 * to perform these two clear_bit() 6579 */ 6580 clear_bit(NAPI_STATE_MISSED, &napi->state); 6581 clear_bit(NAPI_STATE_IN_BUSY_POLL, &napi->state); 6582 6583 local_bh_disable(); 6584 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 6585 6586 if (flags & NAPI_F_PREFER_BUSY_POLL) { 6587 napi->defer_hard_irqs_count = napi_get_defer_hard_irqs(napi); 6588 timeout = napi_get_gro_flush_timeout(napi); 6589 if (napi->defer_hard_irqs_count && timeout) { 6590 hrtimer_start(&napi->timer, ns_to_ktime(timeout), HRTIMER_MODE_REL_PINNED); 6591 skip_schedule = true; 6592 } 6593 } 6594 6595 /* All we really want here is to re-enable device interrupts. 6596 * Ideally, a new ndo_busy_poll_stop() could avoid another round. 6597 */ 6598 rc = napi->poll(napi, budget); 6599 /* We can't gro_normal_list() here, because napi->poll() might have 6600 * rearmed the napi (napi_complete_done()) in which case it could 6601 * already be running on another CPU. 6602 */ 6603 trace_napi_poll(napi, rc, budget); 6604 netpoll_poll_unlock(have_poll_lock); 6605 if (rc == budget) 6606 __busy_poll_stop(napi, skip_schedule); 6607 bpf_net_ctx_clear(bpf_net_ctx); 6608 local_bh_enable(); 6609 } 6610 6611 static void __napi_busy_loop(unsigned int napi_id, 6612 bool (*loop_end)(void *, unsigned long), 6613 void *loop_end_arg, unsigned flags, u16 budget) 6614 { 6615 unsigned long start_time = loop_end ? busy_loop_current_time() : 0; 6616 int (*napi_poll)(struct napi_struct *napi, int budget); 6617 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 6618 void *have_poll_lock = NULL; 6619 struct napi_struct *napi; 6620 6621 WARN_ON_ONCE(!rcu_read_lock_held()); 6622 6623 restart: 6624 napi_poll = NULL; 6625 6626 napi = napi_by_id(napi_id); 6627 if (!napi) 6628 return; 6629 6630 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6631 preempt_disable(); 6632 for (;;) { 6633 int work = 0; 6634 6635 local_bh_disable(); 6636 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 6637 if (!napi_poll) { 6638 unsigned long val = READ_ONCE(napi->state); 6639 6640 /* If multiple threads are competing for this napi, 6641 * we avoid dirtying napi->state as much as we can. 6642 */ 6643 if (val & (NAPIF_STATE_DISABLE | NAPIF_STATE_SCHED | 6644 NAPIF_STATE_IN_BUSY_POLL)) { 6645 if (flags & NAPI_F_PREFER_BUSY_POLL) 6646 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6647 goto count; 6648 } 6649 if (cmpxchg(&napi->state, val, 6650 val | NAPIF_STATE_IN_BUSY_POLL | 6651 NAPIF_STATE_SCHED) != val) { 6652 if (flags & NAPI_F_PREFER_BUSY_POLL) 6653 set_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6654 goto count; 6655 } 6656 have_poll_lock = netpoll_poll_lock(napi); 6657 napi_poll = napi->poll; 6658 } 6659 work = napi_poll(napi, budget); 6660 trace_napi_poll(napi, work, budget); 6661 gro_normal_list(&napi->gro); 6662 count: 6663 if (work > 0) 6664 __NET_ADD_STATS(dev_net(napi->dev), 6665 LINUX_MIB_BUSYPOLLRXPACKETS, work); 6666 skb_defer_free_flush(this_cpu_ptr(&softnet_data)); 6667 bpf_net_ctx_clear(bpf_net_ctx); 6668 local_bh_enable(); 6669 6670 if (!loop_end || loop_end(loop_end_arg, start_time)) 6671 break; 6672 6673 if (unlikely(need_resched())) { 6674 if (flags & NAPI_F_END_ON_RESCHED) 6675 break; 6676 if (napi_poll) 6677 busy_poll_stop(napi, have_poll_lock, flags, budget); 6678 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6679 preempt_enable(); 6680 rcu_read_unlock(); 6681 cond_resched(); 6682 rcu_read_lock(); 6683 if (loop_end(loop_end_arg, start_time)) 6684 return; 6685 goto restart; 6686 } 6687 cpu_relax(); 6688 } 6689 if (napi_poll) 6690 busy_poll_stop(napi, have_poll_lock, flags, budget); 6691 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) 6692 preempt_enable(); 6693 } 6694 6695 void napi_busy_loop_rcu(unsigned int napi_id, 6696 bool (*loop_end)(void *, unsigned long), 6697 void *loop_end_arg, bool prefer_busy_poll, u16 budget) 6698 { 6699 unsigned flags = NAPI_F_END_ON_RESCHED; 6700 6701 if (prefer_busy_poll) 6702 flags |= NAPI_F_PREFER_BUSY_POLL; 6703 6704 __napi_busy_loop(napi_id, loop_end, loop_end_arg, flags, budget); 6705 } 6706 6707 void napi_busy_loop(unsigned int napi_id, 6708 bool (*loop_end)(void *, unsigned long), 6709 void *loop_end_arg, bool prefer_busy_poll, u16 budget) 6710 { 6711 unsigned flags = prefer_busy_poll ? NAPI_F_PREFER_BUSY_POLL : 0; 6712 6713 rcu_read_lock(); 6714 __napi_busy_loop(napi_id, loop_end, loop_end_arg, flags, budget); 6715 rcu_read_unlock(); 6716 } 6717 EXPORT_SYMBOL(napi_busy_loop); 6718 6719 void napi_suspend_irqs(unsigned int napi_id) 6720 { 6721 struct napi_struct *napi; 6722 6723 rcu_read_lock(); 6724 napi = napi_by_id(napi_id); 6725 if (napi) { 6726 unsigned long timeout = napi_get_irq_suspend_timeout(napi); 6727 6728 if (timeout) 6729 hrtimer_start(&napi->timer, ns_to_ktime(timeout), 6730 HRTIMER_MODE_REL_PINNED); 6731 } 6732 rcu_read_unlock(); 6733 } 6734 6735 void napi_resume_irqs(unsigned int napi_id) 6736 { 6737 struct napi_struct *napi; 6738 6739 rcu_read_lock(); 6740 napi = napi_by_id(napi_id); 6741 if (napi) { 6742 /* If irq_suspend_timeout is set to 0 between the call to 6743 * napi_suspend_irqs and now, the original value still 6744 * determines the safety timeout as intended and napi_watchdog 6745 * will resume irq processing. 6746 */ 6747 if (napi_get_irq_suspend_timeout(napi)) { 6748 local_bh_disable(); 6749 napi_schedule(napi); 6750 local_bh_enable(); 6751 } 6752 } 6753 rcu_read_unlock(); 6754 } 6755 6756 #endif /* CONFIG_NET_RX_BUSY_POLL */ 6757 6758 static void __napi_hash_add_with_id(struct napi_struct *napi, 6759 unsigned int napi_id) 6760 { 6761 napi->gro.cached_napi_id = napi_id; 6762 6763 WRITE_ONCE(napi->napi_id, napi_id); 6764 hlist_add_head_rcu(&napi->napi_hash_node, 6765 &napi_hash[napi->napi_id % HASH_SIZE(napi_hash)]); 6766 } 6767 6768 static void napi_hash_add_with_id(struct napi_struct *napi, 6769 unsigned int napi_id) 6770 { 6771 unsigned long flags; 6772 6773 spin_lock_irqsave(&napi_hash_lock, flags); 6774 WARN_ON_ONCE(napi_by_id(napi_id)); 6775 __napi_hash_add_with_id(napi, napi_id); 6776 spin_unlock_irqrestore(&napi_hash_lock, flags); 6777 } 6778 6779 static void napi_hash_add(struct napi_struct *napi) 6780 { 6781 unsigned long flags; 6782 6783 if (test_bit(NAPI_STATE_NO_BUSY_POLL, &napi->state)) 6784 return; 6785 6786 spin_lock_irqsave(&napi_hash_lock, flags); 6787 6788 /* 0..NR_CPUS range is reserved for sender_cpu use */ 6789 do { 6790 if (unlikely(!napi_id_valid(++napi_gen_id))) 6791 napi_gen_id = MIN_NAPI_ID; 6792 } while (napi_by_id(napi_gen_id)); 6793 6794 __napi_hash_add_with_id(napi, napi_gen_id); 6795 6796 spin_unlock_irqrestore(&napi_hash_lock, flags); 6797 } 6798 6799 /* Warning : caller is responsible to make sure rcu grace period 6800 * is respected before freeing memory containing @napi 6801 */ 6802 static void napi_hash_del(struct napi_struct *napi) 6803 { 6804 unsigned long flags; 6805 6806 spin_lock_irqsave(&napi_hash_lock, flags); 6807 6808 hlist_del_init_rcu(&napi->napi_hash_node); 6809 6810 spin_unlock_irqrestore(&napi_hash_lock, flags); 6811 } 6812 6813 static enum hrtimer_restart napi_watchdog(struct hrtimer *timer) 6814 { 6815 struct napi_struct *napi; 6816 6817 napi = container_of(timer, struct napi_struct, timer); 6818 6819 /* Note : we use a relaxed variant of napi_schedule_prep() not setting 6820 * NAPI_STATE_MISSED, since we do not react to a device IRQ. 6821 */ 6822 if (!napi_disable_pending(napi) && 6823 !test_and_set_bit(NAPI_STATE_SCHED, &napi->state)) { 6824 clear_bit(NAPI_STATE_PREFER_BUSY_POLL, &napi->state); 6825 __napi_schedule_irqoff(napi); 6826 } 6827 6828 return HRTIMER_NORESTART; 6829 } 6830 6831 int dev_set_threaded(struct net_device *dev, bool threaded) 6832 { 6833 struct napi_struct *napi; 6834 int err = 0; 6835 6836 netdev_assert_locked_or_invisible(dev); 6837 6838 if (dev->threaded == threaded) 6839 return 0; 6840 6841 if (threaded) { 6842 list_for_each_entry(napi, &dev->napi_list, dev_list) { 6843 if (!napi->thread) { 6844 err = napi_kthread_create(napi); 6845 if (err) { 6846 threaded = false; 6847 break; 6848 } 6849 } 6850 } 6851 } 6852 6853 WRITE_ONCE(dev->threaded, threaded); 6854 6855 /* Make sure kthread is created before THREADED bit 6856 * is set. 6857 */ 6858 smp_mb__before_atomic(); 6859 6860 /* Setting/unsetting threaded mode on a napi might not immediately 6861 * take effect, if the current napi instance is actively being 6862 * polled. In this case, the switch between threaded mode and 6863 * softirq mode will happen in the next round of napi_schedule(). 6864 * This should not cause hiccups/stalls to the live traffic. 6865 */ 6866 list_for_each_entry(napi, &dev->napi_list, dev_list) 6867 assign_bit(NAPI_STATE_THREADED, &napi->state, threaded); 6868 6869 return err; 6870 } 6871 EXPORT_SYMBOL(dev_set_threaded); 6872 6873 /** 6874 * netif_queue_set_napi - Associate queue with the napi 6875 * @dev: device to which NAPI and queue belong 6876 * @queue_index: Index of queue 6877 * @type: queue type as RX or TX 6878 * @napi: NAPI context, pass NULL to clear previously set NAPI 6879 * 6880 * Set queue with its corresponding napi context. This should be done after 6881 * registering the NAPI handler for the queue-vector and the queues have been 6882 * mapped to the corresponding interrupt vector. 6883 */ 6884 void netif_queue_set_napi(struct net_device *dev, unsigned int queue_index, 6885 enum netdev_queue_type type, struct napi_struct *napi) 6886 { 6887 struct netdev_rx_queue *rxq; 6888 struct netdev_queue *txq; 6889 6890 if (WARN_ON_ONCE(napi && !napi->dev)) 6891 return; 6892 if (dev->reg_state >= NETREG_REGISTERED) 6893 ASSERT_RTNL(); 6894 6895 switch (type) { 6896 case NETDEV_QUEUE_TYPE_RX: 6897 rxq = __netif_get_rx_queue(dev, queue_index); 6898 rxq->napi = napi; 6899 return; 6900 case NETDEV_QUEUE_TYPE_TX: 6901 txq = netdev_get_tx_queue(dev, queue_index); 6902 txq->napi = napi; 6903 return; 6904 default: 6905 return; 6906 } 6907 } 6908 EXPORT_SYMBOL(netif_queue_set_napi); 6909 6910 static void 6911 netif_napi_irq_notify(struct irq_affinity_notify *notify, 6912 const cpumask_t *mask) 6913 { 6914 struct napi_struct *napi = 6915 container_of(notify, struct napi_struct, notify); 6916 #ifdef CONFIG_RFS_ACCEL 6917 struct cpu_rmap *rmap = napi->dev->rx_cpu_rmap; 6918 int err; 6919 #endif 6920 6921 if (napi->config && napi->dev->irq_affinity_auto) 6922 cpumask_copy(&napi->config->affinity_mask, mask); 6923 6924 #ifdef CONFIG_RFS_ACCEL 6925 if (napi->dev->rx_cpu_rmap_auto) { 6926 err = cpu_rmap_update(rmap, napi->napi_rmap_idx, mask); 6927 if (err) 6928 netdev_warn(napi->dev, "RMAP update failed (%d)\n", 6929 err); 6930 } 6931 #endif 6932 } 6933 6934 #ifdef CONFIG_RFS_ACCEL 6935 static void netif_napi_affinity_release(struct kref *ref) 6936 { 6937 struct napi_struct *napi = 6938 container_of(ref, struct napi_struct, notify.kref); 6939 struct cpu_rmap *rmap = napi->dev->rx_cpu_rmap; 6940 6941 netdev_assert_locked(napi->dev); 6942 WARN_ON(test_and_clear_bit(NAPI_STATE_HAS_NOTIFIER, 6943 &napi->state)); 6944 6945 if (!napi->dev->rx_cpu_rmap_auto) 6946 return; 6947 rmap->obj[napi->napi_rmap_idx] = NULL; 6948 napi->napi_rmap_idx = -1; 6949 cpu_rmap_put(rmap); 6950 } 6951 6952 int netif_enable_cpu_rmap(struct net_device *dev, unsigned int num_irqs) 6953 { 6954 if (dev->rx_cpu_rmap_auto) 6955 return 0; 6956 6957 dev->rx_cpu_rmap = alloc_irq_cpu_rmap(num_irqs); 6958 if (!dev->rx_cpu_rmap) 6959 return -ENOMEM; 6960 6961 dev->rx_cpu_rmap_auto = true; 6962 return 0; 6963 } 6964 EXPORT_SYMBOL(netif_enable_cpu_rmap); 6965 6966 static void netif_del_cpu_rmap(struct net_device *dev) 6967 { 6968 struct cpu_rmap *rmap = dev->rx_cpu_rmap; 6969 6970 if (!dev->rx_cpu_rmap_auto) 6971 return; 6972 6973 /* Free the rmap */ 6974 cpu_rmap_put(rmap); 6975 dev->rx_cpu_rmap = NULL; 6976 dev->rx_cpu_rmap_auto = false; 6977 } 6978 6979 #else 6980 static void netif_napi_affinity_release(struct kref *ref) 6981 { 6982 } 6983 6984 int netif_enable_cpu_rmap(struct net_device *dev, unsigned int num_irqs) 6985 { 6986 return 0; 6987 } 6988 EXPORT_SYMBOL(netif_enable_cpu_rmap); 6989 6990 static void netif_del_cpu_rmap(struct net_device *dev) 6991 { 6992 } 6993 #endif 6994 6995 void netif_set_affinity_auto(struct net_device *dev) 6996 { 6997 unsigned int i, maxqs, numa; 6998 6999 maxqs = max(dev->num_tx_queues, dev->num_rx_queues); 7000 numa = dev_to_node(&dev->dev); 7001 7002 for (i = 0; i < maxqs; i++) 7003 cpumask_set_cpu(cpumask_local_spread(i, numa), 7004 &dev->napi_config[i].affinity_mask); 7005 7006 dev->irq_affinity_auto = true; 7007 } 7008 EXPORT_SYMBOL(netif_set_affinity_auto); 7009 7010 void netif_napi_set_irq_locked(struct napi_struct *napi, int irq) 7011 { 7012 int rc; 7013 7014 netdev_assert_locked_or_invisible(napi->dev); 7015 7016 if (napi->irq == irq) 7017 return; 7018 7019 /* Remove existing resources */ 7020 if (test_and_clear_bit(NAPI_STATE_HAS_NOTIFIER, &napi->state)) 7021 irq_set_affinity_notifier(napi->irq, NULL); 7022 7023 napi->irq = irq; 7024 if (irq < 0 || 7025 (!napi->dev->rx_cpu_rmap_auto && !napi->dev->irq_affinity_auto)) 7026 return; 7027 7028 /* Abort for buggy drivers */ 7029 if (napi->dev->irq_affinity_auto && WARN_ON_ONCE(!napi->config)) 7030 return; 7031 7032 #ifdef CONFIG_RFS_ACCEL 7033 if (napi->dev->rx_cpu_rmap_auto) { 7034 rc = cpu_rmap_add(napi->dev->rx_cpu_rmap, napi); 7035 if (rc < 0) 7036 return; 7037 7038 cpu_rmap_get(napi->dev->rx_cpu_rmap); 7039 napi->napi_rmap_idx = rc; 7040 } 7041 #endif 7042 7043 /* Use core IRQ notifier */ 7044 napi->notify.notify = netif_napi_irq_notify; 7045 napi->notify.release = netif_napi_affinity_release; 7046 rc = irq_set_affinity_notifier(irq, &napi->notify); 7047 if (rc) { 7048 netdev_warn(napi->dev, "Unable to set IRQ notifier (%d)\n", 7049 rc); 7050 goto put_rmap; 7051 } 7052 7053 set_bit(NAPI_STATE_HAS_NOTIFIER, &napi->state); 7054 return; 7055 7056 put_rmap: 7057 #ifdef CONFIG_RFS_ACCEL 7058 if (napi->dev->rx_cpu_rmap_auto) { 7059 napi->dev->rx_cpu_rmap->obj[napi->napi_rmap_idx] = NULL; 7060 cpu_rmap_put(napi->dev->rx_cpu_rmap); 7061 napi->napi_rmap_idx = -1; 7062 } 7063 #endif 7064 napi->notify.notify = NULL; 7065 napi->notify.release = NULL; 7066 } 7067 EXPORT_SYMBOL(netif_napi_set_irq_locked); 7068 7069 static void napi_restore_config(struct napi_struct *n) 7070 { 7071 n->defer_hard_irqs = n->config->defer_hard_irqs; 7072 n->gro_flush_timeout = n->config->gro_flush_timeout; 7073 n->irq_suspend_timeout = n->config->irq_suspend_timeout; 7074 7075 if (n->dev->irq_affinity_auto && 7076 test_bit(NAPI_STATE_HAS_NOTIFIER, &n->state)) 7077 irq_set_affinity(n->irq, &n->config->affinity_mask); 7078 7079 /* a NAPI ID might be stored in the config, if so use it. if not, use 7080 * napi_hash_add to generate one for us. 7081 */ 7082 if (n->config->napi_id) { 7083 napi_hash_add_with_id(n, n->config->napi_id); 7084 } else { 7085 napi_hash_add(n); 7086 n->config->napi_id = n->napi_id; 7087 } 7088 } 7089 7090 static void napi_save_config(struct napi_struct *n) 7091 { 7092 n->config->defer_hard_irqs = n->defer_hard_irqs; 7093 n->config->gro_flush_timeout = n->gro_flush_timeout; 7094 n->config->irq_suspend_timeout = n->irq_suspend_timeout; 7095 napi_hash_del(n); 7096 } 7097 7098 /* Netlink wants the NAPI list to be sorted by ID, if adding a NAPI which will 7099 * inherit an existing ID try to insert it at the right position. 7100 */ 7101 static void 7102 netif_napi_dev_list_add(struct net_device *dev, struct napi_struct *napi) 7103 { 7104 unsigned int new_id, pos_id; 7105 struct list_head *higher; 7106 struct napi_struct *pos; 7107 7108 new_id = UINT_MAX; 7109 if (napi->config && napi->config->napi_id) 7110 new_id = napi->config->napi_id; 7111 7112 higher = &dev->napi_list; 7113 list_for_each_entry(pos, &dev->napi_list, dev_list) { 7114 if (napi_id_valid(pos->napi_id)) 7115 pos_id = pos->napi_id; 7116 else if (pos->config) 7117 pos_id = pos->config->napi_id; 7118 else 7119 pos_id = UINT_MAX; 7120 7121 if (pos_id <= new_id) 7122 break; 7123 higher = &pos->dev_list; 7124 } 7125 list_add_rcu(&napi->dev_list, higher); /* adds after higher */ 7126 } 7127 7128 /* Double check that napi_get_frags() allocates skbs with 7129 * skb->head being backed by slab, not a page fragment. 7130 * This is to make sure bug fixed in 3226b158e67c 7131 * ("net: avoid 32 x truesize under-estimation for tiny skbs") 7132 * does not accidentally come back. 7133 */ 7134 static void napi_get_frags_check(struct napi_struct *napi) 7135 { 7136 struct sk_buff *skb; 7137 7138 local_bh_disable(); 7139 skb = napi_get_frags(napi); 7140 WARN_ON_ONCE(skb && skb->head_frag); 7141 napi_free_frags(napi); 7142 local_bh_enable(); 7143 } 7144 7145 void netif_napi_add_weight_locked(struct net_device *dev, 7146 struct napi_struct *napi, 7147 int (*poll)(struct napi_struct *, int), 7148 int weight) 7149 { 7150 netdev_assert_locked(dev); 7151 if (WARN_ON(test_and_set_bit(NAPI_STATE_LISTED, &napi->state))) 7152 return; 7153 7154 INIT_LIST_HEAD(&napi->poll_list); 7155 INIT_HLIST_NODE(&napi->napi_hash_node); 7156 hrtimer_init(&napi->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED); 7157 napi->timer.function = napi_watchdog; 7158 gro_init(&napi->gro); 7159 napi->skb = NULL; 7160 napi->poll = poll; 7161 if (weight > NAPI_POLL_WEIGHT) 7162 netdev_err_once(dev, "%s() called with weight %d\n", __func__, 7163 weight); 7164 napi->weight = weight; 7165 napi->dev = dev; 7166 #ifdef CONFIG_NETPOLL 7167 napi->poll_owner = -1; 7168 #endif 7169 napi->list_owner = -1; 7170 set_bit(NAPI_STATE_SCHED, &napi->state); 7171 set_bit(NAPI_STATE_NPSVC, &napi->state); 7172 netif_napi_dev_list_add(dev, napi); 7173 7174 /* default settings from sysfs are applied to all NAPIs. any per-NAPI 7175 * configuration will be loaded in napi_enable 7176 */ 7177 napi_set_defer_hard_irqs(napi, READ_ONCE(dev->napi_defer_hard_irqs)); 7178 napi_set_gro_flush_timeout(napi, READ_ONCE(dev->gro_flush_timeout)); 7179 7180 napi_get_frags_check(napi); 7181 /* Create kthread for this napi if dev->threaded is set. 7182 * Clear dev->threaded if kthread creation failed so that 7183 * threaded mode will not be enabled in napi_enable(). 7184 */ 7185 if (dev->threaded && napi_kthread_create(napi)) 7186 dev->threaded = false; 7187 netif_napi_set_irq_locked(napi, -1); 7188 } 7189 EXPORT_SYMBOL(netif_napi_add_weight_locked); 7190 7191 void napi_disable_locked(struct napi_struct *n) 7192 { 7193 unsigned long val, new; 7194 7195 might_sleep(); 7196 netdev_assert_locked(n->dev); 7197 7198 set_bit(NAPI_STATE_DISABLE, &n->state); 7199 7200 val = READ_ONCE(n->state); 7201 do { 7202 while (val & (NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC)) { 7203 usleep_range(20, 200); 7204 val = READ_ONCE(n->state); 7205 } 7206 7207 new = val | NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC; 7208 new &= ~(NAPIF_STATE_THREADED | NAPIF_STATE_PREFER_BUSY_POLL); 7209 } while (!try_cmpxchg(&n->state, &val, new)); 7210 7211 hrtimer_cancel(&n->timer); 7212 7213 if (n->config) 7214 napi_save_config(n); 7215 else 7216 napi_hash_del(n); 7217 7218 clear_bit(NAPI_STATE_DISABLE, &n->state); 7219 } 7220 EXPORT_SYMBOL(napi_disable_locked); 7221 7222 /** 7223 * napi_disable() - prevent NAPI from scheduling 7224 * @n: NAPI context 7225 * 7226 * Stop NAPI from being scheduled on this context. 7227 * Waits till any outstanding processing completes. 7228 * Takes netdev_lock() for associated net_device. 7229 */ 7230 void napi_disable(struct napi_struct *n) 7231 { 7232 netdev_lock(n->dev); 7233 napi_disable_locked(n); 7234 netdev_unlock(n->dev); 7235 } 7236 EXPORT_SYMBOL(napi_disable); 7237 7238 void napi_enable_locked(struct napi_struct *n) 7239 { 7240 unsigned long new, val = READ_ONCE(n->state); 7241 7242 if (n->config) 7243 napi_restore_config(n); 7244 else 7245 napi_hash_add(n); 7246 7247 do { 7248 BUG_ON(!test_bit(NAPI_STATE_SCHED, &val)); 7249 7250 new = val & ~(NAPIF_STATE_SCHED | NAPIF_STATE_NPSVC); 7251 if (n->dev->threaded && n->thread) 7252 new |= NAPIF_STATE_THREADED; 7253 } while (!try_cmpxchg(&n->state, &val, new)); 7254 } 7255 EXPORT_SYMBOL(napi_enable_locked); 7256 7257 /** 7258 * napi_enable() - enable NAPI scheduling 7259 * @n: NAPI context 7260 * 7261 * Enable scheduling of a NAPI instance. 7262 * Must be paired with napi_disable(). 7263 * Takes netdev_lock() for associated net_device. 7264 */ 7265 void napi_enable(struct napi_struct *n) 7266 { 7267 netdev_lock(n->dev); 7268 napi_enable_locked(n); 7269 netdev_unlock(n->dev); 7270 } 7271 EXPORT_SYMBOL(napi_enable); 7272 7273 /* Must be called in process context */ 7274 void __netif_napi_del_locked(struct napi_struct *napi) 7275 { 7276 netdev_assert_locked(napi->dev); 7277 7278 if (!test_and_clear_bit(NAPI_STATE_LISTED, &napi->state)) 7279 return; 7280 7281 /* Make sure NAPI is disabled (or was never enabled). */ 7282 WARN_ON(!test_bit(NAPI_STATE_SCHED, &napi->state)); 7283 7284 if (test_and_clear_bit(NAPI_STATE_HAS_NOTIFIER, &napi->state)) 7285 irq_set_affinity_notifier(napi->irq, NULL); 7286 7287 if (napi->config) { 7288 napi->index = -1; 7289 napi->config = NULL; 7290 } 7291 7292 list_del_rcu(&napi->dev_list); 7293 napi_free_frags(napi); 7294 7295 gro_cleanup(&napi->gro); 7296 7297 if (napi->thread) { 7298 kthread_stop(napi->thread); 7299 napi->thread = NULL; 7300 } 7301 } 7302 EXPORT_SYMBOL(__netif_napi_del_locked); 7303 7304 static int __napi_poll(struct napi_struct *n, bool *repoll) 7305 { 7306 int work, weight; 7307 7308 weight = n->weight; 7309 7310 /* This NAPI_STATE_SCHED test is for avoiding a race 7311 * with netpoll's poll_napi(). Only the entity which 7312 * obtains the lock and sees NAPI_STATE_SCHED set will 7313 * actually make the ->poll() call. Therefore we avoid 7314 * accidentally calling ->poll() when NAPI is not scheduled. 7315 */ 7316 work = 0; 7317 if (napi_is_scheduled(n)) { 7318 work = n->poll(n, weight); 7319 trace_napi_poll(n, work, weight); 7320 7321 xdp_do_check_flushed(n); 7322 } 7323 7324 if (unlikely(work > weight)) 7325 netdev_err_once(n->dev, "NAPI poll function %pS returned %d, exceeding its budget of %d.\n", 7326 n->poll, work, weight); 7327 7328 if (likely(work < weight)) 7329 return work; 7330 7331 /* Drivers must not modify the NAPI state if they 7332 * consume the entire weight. In such cases this code 7333 * still "owns" the NAPI instance and therefore can 7334 * move the instance around on the list at-will. 7335 */ 7336 if (unlikely(napi_disable_pending(n))) { 7337 napi_complete(n); 7338 return work; 7339 } 7340 7341 /* The NAPI context has more processing work, but busy-polling 7342 * is preferred. Exit early. 7343 */ 7344 if (napi_prefer_busy_poll(n)) { 7345 if (napi_complete_done(n, work)) { 7346 /* If timeout is not set, we need to make sure 7347 * that the NAPI is re-scheduled. 7348 */ 7349 napi_schedule(n); 7350 } 7351 return work; 7352 } 7353 7354 /* Flush too old packets. If HZ < 1000, flush all packets */ 7355 gro_flush(&n->gro, HZ >= 1000); 7356 gro_normal_list(&n->gro); 7357 7358 /* Some drivers may have called napi_schedule 7359 * prior to exhausting their budget. 7360 */ 7361 if (unlikely(!list_empty(&n->poll_list))) { 7362 pr_warn_once("%s: Budget exhausted after napi rescheduled\n", 7363 n->dev ? n->dev->name : "backlog"); 7364 return work; 7365 } 7366 7367 *repoll = true; 7368 7369 return work; 7370 } 7371 7372 static int napi_poll(struct napi_struct *n, struct list_head *repoll) 7373 { 7374 bool do_repoll = false; 7375 void *have; 7376 int work; 7377 7378 list_del_init(&n->poll_list); 7379 7380 have = netpoll_poll_lock(n); 7381 7382 work = __napi_poll(n, &do_repoll); 7383 7384 if (do_repoll) 7385 list_add_tail(&n->poll_list, repoll); 7386 7387 netpoll_poll_unlock(have); 7388 7389 return work; 7390 } 7391 7392 static int napi_thread_wait(struct napi_struct *napi) 7393 { 7394 set_current_state(TASK_INTERRUPTIBLE); 7395 7396 while (!kthread_should_stop()) { 7397 /* Testing SCHED_THREADED bit here to make sure the current 7398 * kthread owns this napi and could poll on this napi. 7399 * Testing SCHED bit is not enough because SCHED bit might be 7400 * set by some other busy poll thread or by napi_disable(). 7401 */ 7402 if (test_bit(NAPI_STATE_SCHED_THREADED, &napi->state)) { 7403 WARN_ON(!list_empty(&napi->poll_list)); 7404 __set_current_state(TASK_RUNNING); 7405 return 0; 7406 } 7407 7408 schedule(); 7409 set_current_state(TASK_INTERRUPTIBLE); 7410 } 7411 __set_current_state(TASK_RUNNING); 7412 7413 return -1; 7414 } 7415 7416 static void napi_threaded_poll_loop(struct napi_struct *napi) 7417 { 7418 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 7419 struct softnet_data *sd; 7420 unsigned long last_qs = jiffies; 7421 7422 for (;;) { 7423 bool repoll = false; 7424 void *have; 7425 7426 local_bh_disable(); 7427 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 7428 7429 sd = this_cpu_ptr(&softnet_data); 7430 sd->in_napi_threaded_poll = true; 7431 7432 have = netpoll_poll_lock(napi); 7433 __napi_poll(napi, &repoll); 7434 netpoll_poll_unlock(have); 7435 7436 sd->in_napi_threaded_poll = false; 7437 barrier(); 7438 7439 if (sd_has_rps_ipi_waiting(sd)) { 7440 local_irq_disable(); 7441 net_rps_action_and_irq_enable(sd); 7442 } 7443 skb_defer_free_flush(sd); 7444 bpf_net_ctx_clear(bpf_net_ctx); 7445 local_bh_enable(); 7446 7447 if (!repoll) 7448 break; 7449 7450 rcu_softirq_qs_periodic(last_qs); 7451 cond_resched(); 7452 } 7453 } 7454 7455 static int napi_threaded_poll(void *data) 7456 { 7457 struct napi_struct *napi = data; 7458 7459 while (!napi_thread_wait(napi)) 7460 napi_threaded_poll_loop(napi); 7461 7462 return 0; 7463 } 7464 7465 static __latent_entropy void net_rx_action(void) 7466 { 7467 struct softnet_data *sd = this_cpu_ptr(&softnet_data); 7468 unsigned long time_limit = jiffies + 7469 usecs_to_jiffies(READ_ONCE(net_hotdata.netdev_budget_usecs)); 7470 struct bpf_net_context __bpf_net_ctx, *bpf_net_ctx; 7471 int budget = READ_ONCE(net_hotdata.netdev_budget); 7472 LIST_HEAD(list); 7473 LIST_HEAD(repoll); 7474 7475 bpf_net_ctx = bpf_net_ctx_set(&__bpf_net_ctx); 7476 start: 7477 sd->in_net_rx_action = true; 7478 local_irq_disable(); 7479 list_splice_init(&sd->poll_list, &list); 7480 local_irq_enable(); 7481 7482 for (;;) { 7483 struct napi_struct *n; 7484 7485 skb_defer_free_flush(sd); 7486 7487 if (list_empty(&list)) { 7488 if (list_empty(&repoll)) { 7489 sd->in_net_rx_action = false; 7490 barrier(); 7491 /* We need to check if ____napi_schedule() 7492 * had refilled poll_list while 7493 * sd->in_net_rx_action was true. 7494 */ 7495 if (!list_empty(&sd->poll_list)) 7496 goto start; 7497 if (!sd_has_rps_ipi_waiting(sd)) 7498 goto end; 7499 } 7500 break; 7501 } 7502 7503 n = list_first_entry(&list, struct napi_struct, poll_list); 7504 budget -= napi_poll(n, &repoll); 7505 7506 /* If softirq window is exhausted then punt. 7507 * Allow this to run for 2 jiffies since which will allow 7508 * an average latency of 1.5/HZ. 7509 */ 7510 if (unlikely(budget <= 0 || 7511 time_after_eq(jiffies, time_limit))) { 7512 sd->time_squeeze++; 7513 break; 7514 } 7515 } 7516 7517 local_irq_disable(); 7518 7519 list_splice_tail_init(&sd->poll_list, &list); 7520 list_splice_tail(&repoll, &list); 7521 list_splice(&list, &sd->poll_list); 7522 if (!list_empty(&sd->poll_list)) 7523 __raise_softirq_irqoff(NET_RX_SOFTIRQ); 7524 else 7525 sd->in_net_rx_action = false; 7526 7527 net_rps_action_and_irq_enable(sd); 7528 end: 7529 bpf_net_ctx_clear(bpf_net_ctx); 7530 } 7531 7532 struct netdev_adjacent { 7533 struct net_device *dev; 7534 netdevice_tracker dev_tracker; 7535 7536 /* upper master flag, there can only be one master device per list */ 7537 bool master; 7538 7539 /* lookup ignore flag */ 7540 bool ignore; 7541 7542 /* counter for the number of times this device was added to us */ 7543 u16 ref_nr; 7544 7545 /* private field for the users */ 7546 void *private; 7547 7548 struct list_head list; 7549 struct rcu_head rcu; 7550 }; 7551 7552 static struct netdev_adjacent *__netdev_find_adj(struct net_device *adj_dev, 7553 struct list_head *adj_list) 7554 { 7555 struct netdev_adjacent *adj; 7556 7557 list_for_each_entry(adj, adj_list, list) { 7558 if (adj->dev == adj_dev) 7559 return adj; 7560 } 7561 return NULL; 7562 } 7563 7564 static int ____netdev_has_upper_dev(struct net_device *upper_dev, 7565 struct netdev_nested_priv *priv) 7566 { 7567 struct net_device *dev = (struct net_device *)priv->data; 7568 7569 return upper_dev == dev; 7570 } 7571 7572 /** 7573 * netdev_has_upper_dev - Check if device is linked to an upper device 7574 * @dev: device 7575 * @upper_dev: upper device to check 7576 * 7577 * Find out if a device is linked to specified upper device and return true 7578 * in case it is. Note that this checks only immediate upper device, 7579 * not through a complete stack of devices. The caller must hold the RTNL lock. 7580 */ 7581 bool netdev_has_upper_dev(struct net_device *dev, 7582 struct net_device *upper_dev) 7583 { 7584 struct netdev_nested_priv priv = { 7585 .data = (void *)upper_dev, 7586 }; 7587 7588 ASSERT_RTNL(); 7589 7590 return netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev, 7591 &priv); 7592 } 7593 EXPORT_SYMBOL(netdev_has_upper_dev); 7594 7595 /** 7596 * netdev_has_upper_dev_all_rcu - Check if device is linked to an upper device 7597 * @dev: device 7598 * @upper_dev: upper device to check 7599 * 7600 * Find out if a device is linked to specified upper device and return true 7601 * in case it is. Note that this checks the entire upper device chain. 7602 * The caller must hold rcu lock. 7603 */ 7604 7605 bool netdev_has_upper_dev_all_rcu(struct net_device *dev, 7606 struct net_device *upper_dev) 7607 { 7608 struct netdev_nested_priv priv = { 7609 .data = (void *)upper_dev, 7610 }; 7611 7612 return !!netdev_walk_all_upper_dev_rcu(dev, ____netdev_has_upper_dev, 7613 &priv); 7614 } 7615 EXPORT_SYMBOL(netdev_has_upper_dev_all_rcu); 7616 7617 /** 7618 * netdev_has_any_upper_dev - Check if device is linked to some device 7619 * @dev: device 7620 * 7621 * Find out if a device is linked to an upper device and return true in case 7622 * it is. The caller must hold the RTNL lock. 7623 */ 7624 bool netdev_has_any_upper_dev(struct net_device *dev) 7625 { 7626 ASSERT_RTNL(); 7627 7628 return !list_empty(&dev->adj_list.upper); 7629 } 7630 EXPORT_SYMBOL(netdev_has_any_upper_dev); 7631 7632 /** 7633 * netdev_master_upper_dev_get - Get master upper device 7634 * @dev: device 7635 * 7636 * Find a master upper device and return pointer to it or NULL in case 7637 * it's not there. The caller must hold the RTNL lock. 7638 */ 7639 struct net_device *netdev_master_upper_dev_get(struct net_device *dev) 7640 { 7641 struct netdev_adjacent *upper; 7642 7643 ASSERT_RTNL(); 7644 7645 if (list_empty(&dev->adj_list.upper)) 7646 return NULL; 7647 7648 upper = list_first_entry(&dev->adj_list.upper, 7649 struct netdev_adjacent, list); 7650 if (likely(upper->master)) 7651 return upper->dev; 7652 return NULL; 7653 } 7654 EXPORT_SYMBOL(netdev_master_upper_dev_get); 7655 7656 static struct net_device *__netdev_master_upper_dev_get(struct net_device *dev) 7657 { 7658 struct netdev_adjacent *upper; 7659 7660 ASSERT_RTNL(); 7661 7662 if (list_empty(&dev->adj_list.upper)) 7663 return NULL; 7664 7665 upper = list_first_entry(&dev->adj_list.upper, 7666 struct netdev_adjacent, list); 7667 if (likely(upper->master) && !upper->ignore) 7668 return upper->dev; 7669 return NULL; 7670 } 7671 7672 /** 7673 * netdev_has_any_lower_dev - Check if device is linked to some device 7674 * @dev: device 7675 * 7676 * Find out if a device is linked to a lower device and return true in case 7677 * it is. The caller must hold the RTNL lock. 7678 */ 7679 static bool netdev_has_any_lower_dev(struct net_device *dev) 7680 { 7681 ASSERT_RTNL(); 7682 7683 return !list_empty(&dev->adj_list.lower); 7684 } 7685 7686 void *netdev_adjacent_get_private(struct list_head *adj_list) 7687 { 7688 struct netdev_adjacent *adj; 7689 7690 adj = list_entry(adj_list, struct netdev_adjacent, list); 7691 7692 return adj->private; 7693 } 7694 EXPORT_SYMBOL(netdev_adjacent_get_private); 7695 7696 /** 7697 * netdev_upper_get_next_dev_rcu - Get the next dev from upper list 7698 * @dev: device 7699 * @iter: list_head ** of the current position 7700 * 7701 * Gets the next device from the dev's upper list, starting from iter 7702 * position. The caller must hold RCU read lock. 7703 */ 7704 struct net_device *netdev_upper_get_next_dev_rcu(struct net_device *dev, 7705 struct list_head **iter) 7706 { 7707 struct netdev_adjacent *upper; 7708 7709 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held()); 7710 7711 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 7712 7713 if (&upper->list == &dev->adj_list.upper) 7714 return NULL; 7715 7716 *iter = &upper->list; 7717 7718 return upper->dev; 7719 } 7720 EXPORT_SYMBOL(netdev_upper_get_next_dev_rcu); 7721 7722 static struct net_device *__netdev_next_upper_dev(struct net_device *dev, 7723 struct list_head **iter, 7724 bool *ignore) 7725 { 7726 struct netdev_adjacent *upper; 7727 7728 upper = list_entry((*iter)->next, struct netdev_adjacent, list); 7729 7730 if (&upper->list == &dev->adj_list.upper) 7731 return NULL; 7732 7733 *iter = &upper->list; 7734 *ignore = upper->ignore; 7735 7736 return upper->dev; 7737 } 7738 7739 static struct net_device *netdev_next_upper_dev_rcu(struct net_device *dev, 7740 struct list_head **iter) 7741 { 7742 struct netdev_adjacent *upper; 7743 7744 WARN_ON_ONCE(!rcu_read_lock_held() && !lockdep_rtnl_is_held()); 7745 7746 upper = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 7747 7748 if (&upper->list == &dev->adj_list.upper) 7749 return NULL; 7750 7751 *iter = &upper->list; 7752 7753 return upper->dev; 7754 } 7755 7756 static int __netdev_walk_all_upper_dev(struct net_device *dev, 7757 int (*fn)(struct net_device *dev, 7758 struct netdev_nested_priv *priv), 7759 struct netdev_nested_priv *priv) 7760 { 7761 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7762 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7763 int ret, cur = 0; 7764 bool ignore; 7765 7766 now = dev; 7767 iter = &dev->adj_list.upper; 7768 7769 while (1) { 7770 if (now != dev) { 7771 ret = fn(now, priv); 7772 if (ret) 7773 return ret; 7774 } 7775 7776 next = NULL; 7777 while (1) { 7778 udev = __netdev_next_upper_dev(now, &iter, &ignore); 7779 if (!udev) 7780 break; 7781 if (ignore) 7782 continue; 7783 7784 next = udev; 7785 niter = &udev->adj_list.upper; 7786 dev_stack[cur] = now; 7787 iter_stack[cur++] = iter; 7788 break; 7789 } 7790 7791 if (!next) { 7792 if (!cur) 7793 return 0; 7794 next = dev_stack[--cur]; 7795 niter = iter_stack[cur]; 7796 } 7797 7798 now = next; 7799 iter = niter; 7800 } 7801 7802 return 0; 7803 } 7804 7805 int netdev_walk_all_upper_dev_rcu(struct net_device *dev, 7806 int (*fn)(struct net_device *dev, 7807 struct netdev_nested_priv *priv), 7808 struct netdev_nested_priv *priv) 7809 { 7810 struct net_device *udev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7811 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7812 int ret, cur = 0; 7813 7814 now = dev; 7815 iter = &dev->adj_list.upper; 7816 7817 while (1) { 7818 if (now != dev) { 7819 ret = fn(now, priv); 7820 if (ret) 7821 return ret; 7822 } 7823 7824 next = NULL; 7825 while (1) { 7826 udev = netdev_next_upper_dev_rcu(now, &iter); 7827 if (!udev) 7828 break; 7829 7830 next = udev; 7831 niter = &udev->adj_list.upper; 7832 dev_stack[cur] = now; 7833 iter_stack[cur++] = iter; 7834 break; 7835 } 7836 7837 if (!next) { 7838 if (!cur) 7839 return 0; 7840 next = dev_stack[--cur]; 7841 niter = iter_stack[cur]; 7842 } 7843 7844 now = next; 7845 iter = niter; 7846 } 7847 7848 return 0; 7849 } 7850 EXPORT_SYMBOL_GPL(netdev_walk_all_upper_dev_rcu); 7851 7852 static bool __netdev_has_upper_dev(struct net_device *dev, 7853 struct net_device *upper_dev) 7854 { 7855 struct netdev_nested_priv priv = { 7856 .flags = 0, 7857 .data = (void *)upper_dev, 7858 }; 7859 7860 ASSERT_RTNL(); 7861 7862 return __netdev_walk_all_upper_dev(dev, ____netdev_has_upper_dev, 7863 &priv); 7864 } 7865 7866 /** 7867 * netdev_lower_get_next_private - Get the next ->private from the 7868 * lower neighbour list 7869 * @dev: device 7870 * @iter: list_head ** of the current position 7871 * 7872 * Gets the next netdev_adjacent->private from the dev's lower neighbour 7873 * list, starting from iter position. The caller must hold either hold the 7874 * RTNL lock or its own locking that guarantees that the neighbour lower 7875 * list will remain unchanged. 7876 */ 7877 void *netdev_lower_get_next_private(struct net_device *dev, 7878 struct list_head **iter) 7879 { 7880 struct netdev_adjacent *lower; 7881 7882 lower = list_entry(*iter, struct netdev_adjacent, list); 7883 7884 if (&lower->list == &dev->adj_list.lower) 7885 return NULL; 7886 7887 *iter = lower->list.next; 7888 7889 return lower->private; 7890 } 7891 EXPORT_SYMBOL(netdev_lower_get_next_private); 7892 7893 /** 7894 * netdev_lower_get_next_private_rcu - Get the next ->private from the 7895 * lower neighbour list, RCU 7896 * variant 7897 * @dev: device 7898 * @iter: list_head ** of the current position 7899 * 7900 * Gets the next netdev_adjacent->private from the dev's lower neighbour 7901 * list, starting from iter position. The caller must hold RCU read lock. 7902 */ 7903 void *netdev_lower_get_next_private_rcu(struct net_device *dev, 7904 struct list_head **iter) 7905 { 7906 struct netdev_adjacent *lower; 7907 7908 WARN_ON_ONCE(!rcu_read_lock_held() && !rcu_read_lock_bh_held()); 7909 7910 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 7911 7912 if (&lower->list == &dev->adj_list.lower) 7913 return NULL; 7914 7915 *iter = &lower->list; 7916 7917 return lower->private; 7918 } 7919 EXPORT_SYMBOL(netdev_lower_get_next_private_rcu); 7920 7921 /** 7922 * netdev_lower_get_next - Get the next device from the lower neighbour 7923 * list 7924 * @dev: device 7925 * @iter: list_head ** of the current position 7926 * 7927 * Gets the next netdev_adjacent from the dev's lower neighbour 7928 * list, starting from iter position. The caller must hold RTNL lock or 7929 * its own locking that guarantees that the neighbour lower 7930 * list will remain unchanged. 7931 */ 7932 void *netdev_lower_get_next(struct net_device *dev, struct list_head **iter) 7933 { 7934 struct netdev_adjacent *lower; 7935 7936 lower = list_entry(*iter, struct netdev_adjacent, list); 7937 7938 if (&lower->list == &dev->adj_list.lower) 7939 return NULL; 7940 7941 *iter = lower->list.next; 7942 7943 return lower->dev; 7944 } 7945 EXPORT_SYMBOL(netdev_lower_get_next); 7946 7947 static struct net_device *netdev_next_lower_dev(struct net_device *dev, 7948 struct list_head **iter) 7949 { 7950 struct netdev_adjacent *lower; 7951 7952 lower = list_entry((*iter)->next, struct netdev_adjacent, list); 7953 7954 if (&lower->list == &dev->adj_list.lower) 7955 return NULL; 7956 7957 *iter = &lower->list; 7958 7959 return lower->dev; 7960 } 7961 7962 static struct net_device *__netdev_next_lower_dev(struct net_device *dev, 7963 struct list_head **iter, 7964 bool *ignore) 7965 { 7966 struct netdev_adjacent *lower; 7967 7968 lower = list_entry((*iter)->next, struct netdev_adjacent, list); 7969 7970 if (&lower->list == &dev->adj_list.lower) 7971 return NULL; 7972 7973 *iter = &lower->list; 7974 *ignore = lower->ignore; 7975 7976 return lower->dev; 7977 } 7978 7979 int netdev_walk_all_lower_dev(struct net_device *dev, 7980 int (*fn)(struct net_device *dev, 7981 struct netdev_nested_priv *priv), 7982 struct netdev_nested_priv *priv) 7983 { 7984 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 7985 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 7986 int ret, cur = 0; 7987 7988 now = dev; 7989 iter = &dev->adj_list.lower; 7990 7991 while (1) { 7992 if (now != dev) { 7993 ret = fn(now, priv); 7994 if (ret) 7995 return ret; 7996 } 7997 7998 next = NULL; 7999 while (1) { 8000 ldev = netdev_next_lower_dev(now, &iter); 8001 if (!ldev) 8002 break; 8003 8004 next = ldev; 8005 niter = &ldev->adj_list.lower; 8006 dev_stack[cur] = now; 8007 iter_stack[cur++] = iter; 8008 break; 8009 } 8010 8011 if (!next) { 8012 if (!cur) 8013 return 0; 8014 next = dev_stack[--cur]; 8015 niter = iter_stack[cur]; 8016 } 8017 8018 now = next; 8019 iter = niter; 8020 } 8021 8022 return 0; 8023 } 8024 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev); 8025 8026 static int __netdev_walk_all_lower_dev(struct net_device *dev, 8027 int (*fn)(struct net_device *dev, 8028 struct netdev_nested_priv *priv), 8029 struct netdev_nested_priv *priv) 8030 { 8031 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8032 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8033 int ret, cur = 0; 8034 bool ignore; 8035 8036 now = dev; 8037 iter = &dev->adj_list.lower; 8038 8039 while (1) { 8040 if (now != dev) { 8041 ret = fn(now, priv); 8042 if (ret) 8043 return ret; 8044 } 8045 8046 next = NULL; 8047 while (1) { 8048 ldev = __netdev_next_lower_dev(now, &iter, &ignore); 8049 if (!ldev) 8050 break; 8051 if (ignore) 8052 continue; 8053 8054 next = ldev; 8055 niter = &ldev->adj_list.lower; 8056 dev_stack[cur] = now; 8057 iter_stack[cur++] = iter; 8058 break; 8059 } 8060 8061 if (!next) { 8062 if (!cur) 8063 return 0; 8064 next = dev_stack[--cur]; 8065 niter = iter_stack[cur]; 8066 } 8067 8068 now = next; 8069 iter = niter; 8070 } 8071 8072 return 0; 8073 } 8074 8075 struct net_device *netdev_next_lower_dev_rcu(struct net_device *dev, 8076 struct list_head **iter) 8077 { 8078 struct netdev_adjacent *lower; 8079 8080 lower = list_entry_rcu((*iter)->next, struct netdev_adjacent, list); 8081 if (&lower->list == &dev->adj_list.lower) 8082 return NULL; 8083 8084 *iter = &lower->list; 8085 8086 return lower->dev; 8087 } 8088 EXPORT_SYMBOL(netdev_next_lower_dev_rcu); 8089 8090 static u8 __netdev_upper_depth(struct net_device *dev) 8091 { 8092 struct net_device *udev; 8093 struct list_head *iter; 8094 u8 max_depth = 0; 8095 bool ignore; 8096 8097 for (iter = &dev->adj_list.upper, 8098 udev = __netdev_next_upper_dev(dev, &iter, &ignore); 8099 udev; 8100 udev = __netdev_next_upper_dev(dev, &iter, &ignore)) { 8101 if (ignore) 8102 continue; 8103 if (max_depth < udev->upper_level) 8104 max_depth = udev->upper_level; 8105 } 8106 8107 return max_depth; 8108 } 8109 8110 static u8 __netdev_lower_depth(struct net_device *dev) 8111 { 8112 struct net_device *ldev; 8113 struct list_head *iter; 8114 u8 max_depth = 0; 8115 bool ignore; 8116 8117 for (iter = &dev->adj_list.lower, 8118 ldev = __netdev_next_lower_dev(dev, &iter, &ignore); 8119 ldev; 8120 ldev = __netdev_next_lower_dev(dev, &iter, &ignore)) { 8121 if (ignore) 8122 continue; 8123 if (max_depth < ldev->lower_level) 8124 max_depth = ldev->lower_level; 8125 } 8126 8127 return max_depth; 8128 } 8129 8130 static int __netdev_update_upper_level(struct net_device *dev, 8131 struct netdev_nested_priv *__unused) 8132 { 8133 dev->upper_level = __netdev_upper_depth(dev) + 1; 8134 return 0; 8135 } 8136 8137 #ifdef CONFIG_LOCKDEP 8138 static LIST_HEAD(net_unlink_list); 8139 8140 static void net_unlink_todo(struct net_device *dev) 8141 { 8142 if (list_empty(&dev->unlink_list)) 8143 list_add_tail(&dev->unlink_list, &net_unlink_list); 8144 } 8145 #endif 8146 8147 static int __netdev_update_lower_level(struct net_device *dev, 8148 struct netdev_nested_priv *priv) 8149 { 8150 dev->lower_level = __netdev_lower_depth(dev) + 1; 8151 8152 #ifdef CONFIG_LOCKDEP 8153 if (!priv) 8154 return 0; 8155 8156 if (priv->flags & NESTED_SYNC_IMM) 8157 dev->nested_level = dev->lower_level - 1; 8158 if (priv->flags & NESTED_SYNC_TODO) 8159 net_unlink_todo(dev); 8160 #endif 8161 return 0; 8162 } 8163 8164 int netdev_walk_all_lower_dev_rcu(struct net_device *dev, 8165 int (*fn)(struct net_device *dev, 8166 struct netdev_nested_priv *priv), 8167 struct netdev_nested_priv *priv) 8168 { 8169 struct net_device *ldev, *next, *now, *dev_stack[MAX_NEST_DEV + 1]; 8170 struct list_head *niter, *iter, *iter_stack[MAX_NEST_DEV + 1]; 8171 int ret, cur = 0; 8172 8173 now = dev; 8174 iter = &dev->adj_list.lower; 8175 8176 while (1) { 8177 if (now != dev) { 8178 ret = fn(now, priv); 8179 if (ret) 8180 return ret; 8181 } 8182 8183 next = NULL; 8184 while (1) { 8185 ldev = netdev_next_lower_dev_rcu(now, &iter); 8186 if (!ldev) 8187 break; 8188 8189 next = ldev; 8190 niter = &ldev->adj_list.lower; 8191 dev_stack[cur] = now; 8192 iter_stack[cur++] = iter; 8193 break; 8194 } 8195 8196 if (!next) { 8197 if (!cur) 8198 return 0; 8199 next = dev_stack[--cur]; 8200 niter = iter_stack[cur]; 8201 } 8202 8203 now = next; 8204 iter = niter; 8205 } 8206 8207 return 0; 8208 } 8209 EXPORT_SYMBOL_GPL(netdev_walk_all_lower_dev_rcu); 8210 8211 /** 8212 * netdev_lower_get_first_private_rcu - Get the first ->private from the 8213 * lower neighbour list, RCU 8214 * variant 8215 * @dev: device 8216 * 8217 * Gets the first netdev_adjacent->private from the dev's lower neighbour 8218 * list. The caller must hold RCU read lock. 8219 */ 8220 void *netdev_lower_get_first_private_rcu(struct net_device *dev) 8221 { 8222 struct netdev_adjacent *lower; 8223 8224 lower = list_first_or_null_rcu(&dev->adj_list.lower, 8225 struct netdev_adjacent, list); 8226 if (lower) 8227 return lower->private; 8228 return NULL; 8229 } 8230 EXPORT_SYMBOL(netdev_lower_get_first_private_rcu); 8231 8232 /** 8233 * netdev_master_upper_dev_get_rcu - Get master upper device 8234 * @dev: device 8235 * 8236 * Find a master upper device and return pointer to it or NULL in case 8237 * it's not there. The caller must hold the RCU read lock. 8238 */ 8239 struct net_device *netdev_master_upper_dev_get_rcu(struct net_device *dev) 8240 { 8241 struct netdev_adjacent *upper; 8242 8243 upper = list_first_or_null_rcu(&dev->adj_list.upper, 8244 struct netdev_adjacent, list); 8245 if (upper && likely(upper->master)) 8246 return upper->dev; 8247 return NULL; 8248 } 8249 EXPORT_SYMBOL(netdev_master_upper_dev_get_rcu); 8250 8251 static int netdev_adjacent_sysfs_add(struct net_device *dev, 8252 struct net_device *adj_dev, 8253 struct list_head *dev_list) 8254 { 8255 char linkname[IFNAMSIZ+7]; 8256 8257 sprintf(linkname, dev_list == &dev->adj_list.upper ? 8258 "upper_%s" : "lower_%s", adj_dev->name); 8259 return sysfs_create_link(&(dev->dev.kobj), &(adj_dev->dev.kobj), 8260 linkname); 8261 } 8262 static void netdev_adjacent_sysfs_del(struct net_device *dev, 8263 char *name, 8264 struct list_head *dev_list) 8265 { 8266 char linkname[IFNAMSIZ+7]; 8267 8268 sprintf(linkname, dev_list == &dev->adj_list.upper ? 8269 "upper_%s" : "lower_%s", name); 8270 sysfs_remove_link(&(dev->dev.kobj), linkname); 8271 } 8272 8273 static inline bool netdev_adjacent_is_neigh_list(struct net_device *dev, 8274 struct net_device *adj_dev, 8275 struct list_head *dev_list) 8276 { 8277 return (dev_list == &dev->adj_list.upper || 8278 dev_list == &dev->adj_list.lower) && 8279 net_eq(dev_net(dev), dev_net(adj_dev)); 8280 } 8281 8282 static int __netdev_adjacent_dev_insert(struct net_device *dev, 8283 struct net_device *adj_dev, 8284 struct list_head *dev_list, 8285 void *private, bool master) 8286 { 8287 struct netdev_adjacent *adj; 8288 int ret; 8289 8290 adj = __netdev_find_adj(adj_dev, dev_list); 8291 8292 if (adj) { 8293 adj->ref_nr += 1; 8294 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d\n", 8295 dev->name, adj_dev->name, adj->ref_nr); 8296 8297 return 0; 8298 } 8299 8300 adj = kmalloc(sizeof(*adj), GFP_KERNEL); 8301 if (!adj) 8302 return -ENOMEM; 8303 8304 adj->dev = adj_dev; 8305 adj->master = master; 8306 adj->ref_nr = 1; 8307 adj->private = private; 8308 adj->ignore = false; 8309 netdev_hold(adj_dev, &adj->dev_tracker, GFP_KERNEL); 8310 8311 pr_debug("Insert adjacency: dev %s adj_dev %s adj->ref_nr %d; dev_hold on %s\n", 8312 dev->name, adj_dev->name, adj->ref_nr, adj_dev->name); 8313 8314 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) { 8315 ret = netdev_adjacent_sysfs_add(dev, adj_dev, dev_list); 8316 if (ret) 8317 goto free_adj; 8318 } 8319 8320 /* Ensure that master link is always the first item in list. */ 8321 if (master) { 8322 ret = sysfs_create_link(&(dev->dev.kobj), 8323 &(adj_dev->dev.kobj), "master"); 8324 if (ret) 8325 goto remove_symlinks; 8326 8327 list_add_rcu(&adj->list, dev_list); 8328 } else { 8329 list_add_tail_rcu(&adj->list, dev_list); 8330 } 8331 8332 return 0; 8333 8334 remove_symlinks: 8335 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) 8336 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list); 8337 free_adj: 8338 netdev_put(adj_dev, &adj->dev_tracker); 8339 kfree(adj); 8340 8341 return ret; 8342 } 8343 8344 static void __netdev_adjacent_dev_remove(struct net_device *dev, 8345 struct net_device *adj_dev, 8346 u16 ref_nr, 8347 struct list_head *dev_list) 8348 { 8349 struct netdev_adjacent *adj; 8350 8351 pr_debug("Remove adjacency: dev %s adj_dev %s ref_nr %d\n", 8352 dev->name, adj_dev->name, ref_nr); 8353 8354 adj = __netdev_find_adj(adj_dev, dev_list); 8355 8356 if (!adj) { 8357 pr_err("Adjacency does not exist for device %s from %s\n", 8358 dev->name, adj_dev->name); 8359 WARN_ON(1); 8360 return; 8361 } 8362 8363 if (adj->ref_nr > ref_nr) { 8364 pr_debug("adjacency: %s to %s ref_nr - %d = %d\n", 8365 dev->name, adj_dev->name, ref_nr, 8366 adj->ref_nr - ref_nr); 8367 adj->ref_nr -= ref_nr; 8368 return; 8369 } 8370 8371 if (adj->master) 8372 sysfs_remove_link(&(dev->dev.kobj), "master"); 8373 8374 if (netdev_adjacent_is_neigh_list(dev, adj_dev, dev_list)) 8375 netdev_adjacent_sysfs_del(dev, adj_dev->name, dev_list); 8376 8377 list_del_rcu(&adj->list); 8378 pr_debug("adjacency: dev_put for %s, because link removed from %s to %s\n", 8379 adj_dev->name, dev->name, adj_dev->name); 8380 netdev_put(adj_dev, &adj->dev_tracker); 8381 kfree_rcu(adj, rcu); 8382 } 8383 8384 static int __netdev_adjacent_dev_link_lists(struct net_device *dev, 8385 struct net_device *upper_dev, 8386 struct list_head *up_list, 8387 struct list_head *down_list, 8388 void *private, bool master) 8389 { 8390 int ret; 8391 8392 ret = __netdev_adjacent_dev_insert(dev, upper_dev, up_list, 8393 private, master); 8394 if (ret) 8395 return ret; 8396 8397 ret = __netdev_adjacent_dev_insert(upper_dev, dev, down_list, 8398 private, false); 8399 if (ret) { 8400 __netdev_adjacent_dev_remove(dev, upper_dev, 1, up_list); 8401 return ret; 8402 } 8403 8404 return 0; 8405 } 8406 8407 static void __netdev_adjacent_dev_unlink_lists(struct net_device *dev, 8408 struct net_device *upper_dev, 8409 u16 ref_nr, 8410 struct list_head *up_list, 8411 struct list_head *down_list) 8412 { 8413 __netdev_adjacent_dev_remove(dev, upper_dev, ref_nr, up_list); 8414 __netdev_adjacent_dev_remove(upper_dev, dev, ref_nr, down_list); 8415 } 8416 8417 static int __netdev_adjacent_dev_link_neighbour(struct net_device *dev, 8418 struct net_device *upper_dev, 8419 void *private, bool master) 8420 { 8421 return __netdev_adjacent_dev_link_lists(dev, upper_dev, 8422 &dev->adj_list.upper, 8423 &upper_dev->adj_list.lower, 8424 private, master); 8425 } 8426 8427 static void __netdev_adjacent_dev_unlink_neighbour(struct net_device *dev, 8428 struct net_device *upper_dev) 8429 { 8430 __netdev_adjacent_dev_unlink_lists(dev, upper_dev, 1, 8431 &dev->adj_list.upper, 8432 &upper_dev->adj_list.lower); 8433 } 8434 8435 static int __netdev_upper_dev_link(struct net_device *dev, 8436 struct net_device *upper_dev, bool master, 8437 void *upper_priv, void *upper_info, 8438 struct netdev_nested_priv *priv, 8439 struct netlink_ext_ack *extack) 8440 { 8441 struct netdev_notifier_changeupper_info changeupper_info = { 8442 .info = { 8443 .dev = dev, 8444 .extack = extack, 8445 }, 8446 .upper_dev = upper_dev, 8447 .master = master, 8448 .linking = true, 8449 .upper_info = upper_info, 8450 }; 8451 struct net_device *master_dev; 8452 int ret = 0; 8453 8454 ASSERT_RTNL(); 8455 8456 if (dev == upper_dev) 8457 return -EBUSY; 8458 8459 /* To prevent loops, check if dev is not upper device to upper_dev. */ 8460 if (__netdev_has_upper_dev(upper_dev, dev)) 8461 return -EBUSY; 8462 8463 if ((dev->lower_level + upper_dev->upper_level) > MAX_NEST_DEV) 8464 return -EMLINK; 8465 8466 if (!master) { 8467 if (__netdev_has_upper_dev(dev, upper_dev)) 8468 return -EEXIST; 8469 } else { 8470 master_dev = __netdev_master_upper_dev_get(dev); 8471 if (master_dev) 8472 return master_dev == upper_dev ? -EEXIST : -EBUSY; 8473 } 8474 8475 ret = call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER, 8476 &changeupper_info.info); 8477 ret = notifier_to_errno(ret); 8478 if (ret) 8479 return ret; 8480 8481 ret = __netdev_adjacent_dev_link_neighbour(dev, upper_dev, upper_priv, 8482 master); 8483 if (ret) 8484 return ret; 8485 8486 ret = call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, 8487 &changeupper_info.info); 8488 ret = notifier_to_errno(ret); 8489 if (ret) 8490 goto rollback; 8491 8492 __netdev_update_upper_level(dev, NULL); 8493 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL); 8494 8495 __netdev_update_lower_level(upper_dev, priv); 8496 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level, 8497 priv); 8498 8499 return 0; 8500 8501 rollback: 8502 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev); 8503 8504 return ret; 8505 } 8506 8507 /** 8508 * netdev_upper_dev_link - Add a link to the upper device 8509 * @dev: device 8510 * @upper_dev: new upper device 8511 * @extack: netlink extended ack 8512 * 8513 * Adds a link to device which is upper to this one. The caller must hold 8514 * the RTNL lock. On a failure a negative errno code is returned. 8515 * On success the reference counts are adjusted and the function 8516 * returns zero. 8517 */ 8518 int netdev_upper_dev_link(struct net_device *dev, 8519 struct net_device *upper_dev, 8520 struct netlink_ext_ack *extack) 8521 { 8522 struct netdev_nested_priv priv = { 8523 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 8524 .data = NULL, 8525 }; 8526 8527 return __netdev_upper_dev_link(dev, upper_dev, false, 8528 NULL, NULL, &priv, extack); 8529 } 8530 EXPORT_SYMBOL(netdev_upper_dev_link); 8531 8532 /** 8533 * netdev_master_upper_dev_link - Add a master link to the upper device 8534 * @dev: device 8535 * @upper_dev: new upper device 8536 * @upper_priv: upper device private 8537 * @upper_info: upper info to be passed down via notifier 8538 * @extack: netlink extended ack 8539 * 8540 * Adds a link to device which is upper to this one. In this case, only 8541 * one master upper device can be linked, although other non-master devices 8542 * might be linked as well. The caller must hold the RTNL lock. 8543 * On a failure a negative errno code is returned. On success the reference 8544 * counts are adjusted and the function returns zero. 8545 */ 8546 int netdev_master_upper_dev_link(struct net_device *dev, 8547 struct net_device *upper_dev, 8548 void *upper_priv, void *upper_info, 8549 struct netlink_ext_ack *extack) 8550 { 8551 struct netdev_nested_priv priv = { 8552 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 8553 .data = NULL, 8554 }; 8555 8556 return __netdev_upper_dev_link(dev, upper_dev, true, 8557 upper_priv, upper_info, &priv, extack); 8558 } 8559 EXPORT_SYMBOL(netdev_master_upper_dev_link); 8560 8561 static void __netdev_upper_dev_unlink(struct net_device *dev, 8562 struct net_device *upper_dev, 8563 struct netdev_nested_priv *priv) 8564 { 8565 struct netdev_notifier_changeupper_info changeupper_info = { 8566 .info = { 8567 .dev = dev, 8568 }, 8569 .upper_dev = upper_dev, 8570 .linking = false, 8571 }; 8572 8573 ASSERT_RTNL(); 8574 8575 changeupper_info.master = netdev_master_upper_dev_get(dev) == upper_dev; 8576 8577 call_netdevice_notifiers_info(NETDEV_PRECHANGEUPPER, 8578 &changeupper_info.info); 8579 8580 __netdev_adjacent_dev_unlink_neighbour(dev, upper_dev); 8581 8582 call_netdevice_notifiers_info(NETDEV_CHANGEUPPER, 8583 &changeupper_info.info); 8584 8585 __netdev_update_upper_level(dev, NULL); 8586 __netdev_walk_all_lower_dev(dev, __netdev_update_upper_level, NULL); 8587 8588 __netdev_update_lower_level(upper_dev, priv); 8589 __netdev_walk_all_upper_dev(upper_dev, __netdev_update_lower_level, 8590 priv); 8591 } 8592 8593 /** 8594 * netdev_upper_dev_unlink - Removes a link to upper device 8595 * @dev: device 8596 * @upper_dev: new upper device 8597 * 8598 * Removes a link to device which is upper to this one. The caller must hold 8599 * the RTNL lock. 8600 */ 8601 void netdev_upper_dev_unlink(struct net_device *dev, 8602 struct net_device *upper_dev) 8603 { 8604 struct netdev_nested_priv priv = { 8605 .flags = NESTED_SYNC_TODO, 8606 .data = NULL, 8607 }; 8608 8609 __netdev_upper_dev_unlink(dev, upper_dev, &priv); 8610 } 8611 EXPORT_SYMBOL(netdev_upper_dev_unlink); 8612 8613 static void __netdev_adjacent_dev_set(struct net_device *upper_dev, 8614 struct net_device *lower_dev, 8615 bool val) 8616 { 8617 struct netdev_adjacent *adj; 8618 8619 adj = __netdev_find_adj(lower_dev, &upper_dev->adj_list.lower); 8620 if (adj) 8621 adj->ignore = val; 8622 8623 adj = __netdev_find_adj(upper_dev, &lower_dev->adj_list.upper); 8624 if (adj) 8625 adj->ignore = val; 8626 } 8627 8628 static void netdev_adjacent_dev_disable(struct net_device *upper_dev, 8629 struct net_device *lower_dev) 8630 { 8631 __netdev_adjacent_dev_set(upper_dev, lower_dev, true); 8632 } 8633 8634 static void netdev_adjacent_dev_enable(struct net_device *upper_dev, 8635 struct net_device *lower_dev) 8636 { 8637 __netdev_adjacent_dev_set(upper_dev, lower_dev, false); 8638 } 8639 8640 int netdev_adjacent_change_prepare(struct net_device *old_dev, 8641 struct net_device *new_dev, 8642 struct net_device *dev, 8643 struct netlink_ext_ack *extack) 8644 { 8645 struct netdev_nested_priv priv = { 8646 .flags = 0, 8647 .data = NULL, 8648 }; 8649 int err; 8650 8651 if (!new_dev) 8652 return 0; 8653 8654 if (old_dev && new_dev != old_dev) 8655 netdev_adjacent_dev_disable(dev, old_dev); 8656 err = __netdev_upper_dev_link(new_dev, dev, false, NULL, NULL, &priv, 8657 extack); 8658 if (err) { 8659 if (old_dev && new_dev != old_dev) 8660 netdev_adjacent_dev_enable(dev, old_dev); 8661 return err; 8662 } 8663 8664 return 0; 8665 } 8666 EXPORT_SYMBOL(netdev_adjacent_change_prepare); 8667 8668 void netdev_adjacent_change_commit(struct net_device *old_dev, 8669 struct net_device *new_dev, 8670 struct net_device *dev) 8671 { 8672 struct netdev_nested_priv priv = { 8673 .flags = NESTED_SYNC_IMM | NESTED_SYNC_TODO, 8674 .data = NULL, 8675 }; 8676 8677 if (!new_dev || !old_dev) 8678 return; 8679 8680 if (new_dev == old_dev) 8681 return; 8682 8683 netdev_adjacent_dev_enable(dev, old_dev); 8684 __netdev_upper_dev_unlink(old_dev, dev, &priv); 8685 } 8686 EXPORT_SYMBOL(netdev_adjacent_change_commit); 8687 8688 void netdev_adjacent_change_abort(struct net_device *old_dev, 8689 struct net_device *new_dev, 8690 struct net_device *dev) 8691 { 8692 struct netdev_nested_priv priv = { 8693 .flags = 0, 8694 .data = NULL, 8695 }; 8696 8697 if (!new_dev) 8698 return; 8699 8700 if (old_dev && new_dev != old_dev) 8701 netdev_adjacent_dev_enable(dev, old_dev); 8702 8703 __netdev_upper_dev_unlink(new_dev, dev, &priv); 8704 } 8705 EXPORT_SYMBOL(netdev_adjacent_change_abort); 8706 8707 /** 8708 * netdev_bonding_info_change - Dispatch event about slave change 8709 * @dev: device 8710 * @bonding_info: info to dispatch 8711 * 8712 * Send NETDEV_BONDING_INFO to netdev notifiers with info. 8713 * The caller must hold the RTNL lock. 8714 */ 8715 void netdev_bonding_info_change(struct net_device *dev, 8716 struct netdev_bonding_info *bonding_info) 8717 { 8718 struct netdev_notifier_bonding_info info = { 8719 .info.dev = dev, 8720 }; 8721 8722 memcpy(&info.bonding_info, bonding_info, 8723 sizeof(struct netdev_bonding_info)); 8724 call_netdevice_notifiers_info(NETDEV_BONDING_INFO, 8725 &info.info); 8726 } 8727 EXPORT_SYMBOL(netdev_bonding_info_change); 8728 8729 static int netdev_offload_xstats_enable_l3(struct net_device *dev, 8730 struct netlink_ext_ack *extack) 8731 { 8732 struct netdev_notifier_offload_xstats_info info = { 8733 .info.dev = dev, 8734 .info.extack = extack, 8735 .type = NETDEV_OFFLOAD_XSTATS_TYPE_L3, 8736 }; 8737 int err; 8738 int rc; 8739 8740 dev->offload_xstats_l3 = kzalloc(sizeof(*dev->offload_xstats_l3), 8741 GFP_KERNEL); 8742 if (!dev->offload_xstats_l3) 8743 return -ENOMEM; 8744 8745 rc = call_netdevice_notifiers_info_robust(NETDEV_OFFLOAD_XSTATS_ENABLE, 8746 NETDEV_OFFLOAD_XSTATS_DISABLE, 8747 &info.info); 8748 err = notifier_to_errno(rc); 8749 if (err) 8750 goto free_stats; 8751 8752 return 0; 8753 8754 free_stats: 8755 kfree(dev->offload_xstats_l3); 8756 dev->offload_xstats_l3 = NULL; 8757 return err; 8758 } 8759 8760 int netdev_offload_xstats_enable(struct net_device *dev, 8761 enum netdev_offload_xstats_type type, 8762 struct netlink_ext_ack *extack) 8763 { 8764 ASSERT_RTNL(); 8765 8766 if (netdev_offload_xstats_enabled(dev, type)) 8767 return -EALREADY; 8768 8769 switch (type) { 8770 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 8771 return netdev_offload_xstats_enable_l3(dev, extack); 8772 } 8773 8774 WARN_ON(1); 8775 return -EINVAL; 8776 } 8777 EXPORT_SYMBOL(netdev_offload_xstats_enable); 8778 8779 static void netdev_offload_xstats_disable_l3(struct net_device *dev) 8780 { 8781 struct netdev_notifier_offload_xstats_info info = { 8782 .info.dev = dev, 8783 .type = NETDEV_OFFLOAD_XSTATS_TYPE_L3, 8784 }; 8785 8786 call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_DISABLE, 8787 &info.info); 8788 kfree(dev->offload_xstats_l3); 8789 dev->offload_xstats_l3 = NULL; 8790 } 8791 8792 int netdev_offload_xstats_disable(struct net_device *dev, 8793 enum netdev_offload_xstats_type type) 8794 { 8795 ASSERT_RTNL(); 8796 8797 if (!netdev_offload_xstats_enabled(dev, type)) 8798 return -EALREADY; 8799 8800 switch (type) { 8801 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 8802 netdev_offload_xstats_disable_l3(dev); 8803 return 0; 8804 } 8805 8806 WARN_ON(1); 8807 return -EINVAL; 8808 } 8809 EXPORT_SYMBOL(netdev_offload_xstats_disable); 8810 8811 static void netdev_offload_xstats_disable_all(struct net_device *dev) 8812 { 8813 netdev_offload_xstats_disable(dev, NETDEV_OFFLOAD_XSTATS_TYPE_L3); 8814 } 8815 8816 static struct rtnl_hw_stats64 * 8817 netdev_offload_xstats_get_ptr(const struct net_device *dev, 8818 enum netdev_offload_xstats_type type) 8819 { 8820 switch (type) { 8821 case NETDEV_OFFLOAD_XSTATS_TYPE_L3: 8822 return dev->offload_xstats_l3; 8823 } 8824 8825 WARN_ON(1); 8826 return NULL; 8827 } 8828 8829 bool netdev_offload_xstats_enabled(const struct net_device *dev, 8830 enum netdev_offload_xstats_type type) 8831 { 8832 ASSERT_RTNL(); 8833 8834 return netdev_offload_xstats_get_ptr(dev, type); 8835 } 8836 EXPORT_SYMBOL(netdev_offload_xstats_enabled); 8837 8838 struct netdev_notifier_offload_xstats_ru { 8839 bool used; 8840 }; 8841 8842 struct netdev_notifier_offload_xstats_rd { 8843 struct rtnl_hw_stats64 stats; 8844 bool used; 8845 }; 8846 8847 static void netdev_hw_stats64_add(struct rtnl_hw_stats64 *dest, 8848 const struct rtnl_hw_stats64 *src) 8849 { 8850 dest->rx_packets += src->rx_packets; 8851 dest->tx_packets += src->tx_packets; 8852 dest->rx_bytes += src->rx_bytes; 8853 dest->tx_bytes += src->tx_bytes; 8854 dest->rx_errors += src->rx_errors; 8855 dest->tx_errors += src->tx_errors; 8856 dest->rx_dropped += src->rx_dropped; 8857 dest->tx_dropped += src->tx_dropped; 8858 dest->multicast += src->multicast; 8859 } 8860 8861 static int netdev_offload_xstats_get_used(struct net_device *dev, 8862 enum netdev_offload_xstats_type type, 8863 bool *p_used, 8864 struct netlink_ext_ack *extack) 8865 { 8866 struct netdev_notifier_offload_xstats_ru report_used = {}; 8867 struct netdev_notifier_offload_xstats_info info = { 8868 .info.dev = dev, 8869 .info.extack = extack, 8870 .type = type, 8871 .report_used = &report_used, 8872 }; 8873 int rc; 8874 8875 WARN_ON(!netdev_offload_xstats_enabled(dev, type)); 8876 rc = call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_REPORT_USED, 8877 &info.info); 8878 *p_used = report_used.used; 8879 return notifier_to_errno(rc); 8880 } 8881 8882 static int netdev_offload_xstats_get_stats(struct net_device *dev, 8883 enum netdev_offload_xstats_type type, 8884 struct rtnl_hw_stats64 *p_stats, 8885 bool *p_used, 8886 struct netlink_ext_ack *extack) 8887 { 8888 struct netdev_notifier_offload_xstats_rd report_delta = {}; 8889 struct netdev_notifier_offload_xstats_info info = { 8890 .info.dev = dev, 8891 .info.extack = extack, 8892 .type = type, 8893 .report_delta = &report_delta, 8894 }; 8895 struct rtnl_hw_stats64 *stats; 8896 int rc; 8897 8898 stats = netdev_offload_xstats_get_ptr(dev, type); 8899 if (WARN_ON(!stats)) 8900 return -EINVAL; 8901 8902 rc = call_netdevice_notifiers_info(NETDEV_OFFLOAD_XSTATS_REPORT_DELTA, 8903 &info.info); 8904 8905 /* Cache whatever we got, even if there was an error, otherwise the 8906 * successful stats retrievals would get lost. 8907 */ 8908 netdev_hw_stats64_add(stats, &report_delta.stats); 8909 8910 if (p_stats) 8911 *p_stats = *stats; 8912 *p_used = report_delta.used; 8913 8914 return notifier_to_errno(rc); 8915 } 8916 8917 int netdev_offload_xstats_get(struct net_device *dev, 8918 enum netdev_offload_xstats_type type, 8919 struct rtnl_hw_stats64 *p_stats, bool *p_used, 8920 struct netlink_ext_ack *extack) 8921 { 8922 ASSERT_RTNL(); 8923 8924 if (p_stats) 8925 return netdev_offload_xstats_get_stats(dev, type, p_stats, 8926 p_used, extack); 8927 else 8928 return netdev_offload_xstats_get_used(dev, type, p_used, 8929 extack); 8930 } 8931 EXPORT_SYMBOL(netdev_offload_xstats_get); 8932 8933 void 8934 netdev_offload_xstats_report_delta(struct netdev_notifier_offload_xstats_rd *report_delta, 8935 const struct rtnl_hw_stats64 *stats) 8936 { 8937 report_delta->used = true; 8938 netdev_hw_stats64_add(&report_delta->stats, stats); 8939 } 8940 EXPORT_SYMBOL(netdev_offload_xstats_report_delta); 8941 8942 void 8943 netdev_offload_xstats_report_used(struct netdev_notifier_offload_xstats_ru *report_used) 8944 { 8945 report_used->used = true; 8946 } 8947 EXPORT_SYMBOL(netdev_offload_xstats_report_used); 8948 8949 void netdev_offload_xstats_push_delta(struct net_device *dev, 8950 enum netdev_offload_xstats_type type, 8951 const struct rtnl_hw_stats64 *p_stats) 8952 { 8953 struct rtnl_hw_stats64 *stats; 8954 8955 ASSERT_RTNL(); 8956 8957 stats = netdev_offload_xstats_get_ptr(dev, type); 8958 if (WARN_ON(!stats)) 8959 return; 8960 8961 netdev_hw_stats64_add(stats, p_stats); 8962 } 8963 EXPORT_SYMBOL(netdev_offload_xstats_push_delta); 8964 8965 /** 8966 * netdev_get_xmit_slave - Get the xmit slave of master device 8967 * @dev: device 8968 * @skb: The packet 8969 * @all_slaves: assume all the slaves are active 8970 * 8971 * The reference counters are not incremented so the caller must be 8972 * careful with locks. The caller must hold RCU lock. 8973 * %NULL is returned if no slave is found. 8974 */ 8975 8976 struct net_device *netdev_get_xmit_slave(struct net_device *dev, 8977 struct sk_buff *skb, 8978 bool all_slaves) 8979 { 8980 const struct net_device_ops *ops = dev->netdev_ops; 8981 8982 if (!ops->ndo_get_xmit_slave) 8983 return NULL; 8984 return ops->ndo_get_xmit_slave(dev, skb, all_slaves); 8985 } 8986 EXPORT_SYMBOL(netdev_get_xmit_slave); 8987 8988 static struct net_device *netdev_sk_get_lower_dev(struct net_device *dev, 8989 struct sock *sk) 8990 { 8991 const struct net_device_ops *ops = dev->netdev_ops; 8992 8993 if (!ops->ndo_sk_get_lower_dev) 8994 return NULL; 8995 return ops->ndo_sk_get_lower_dev(dev, sk); 8996 } 8997 8998 /** 8999 * netdev_sk_get_lowest_dev - Get the lowest device in chain given device and socket 9000 * @dev: device 9001 * @sk: the socket 9002 * 9003 * %NULL is returned if no lower device is found. 9004 */ 9005 9006 struct net_device *netdev_sk_get_lowest_dev(struct net_device *dev, 9007 struct sock *sk) 9008 { 9009 struct net_device *lower; 9010 9011 lower = netdev_sk_get_lower_dev(dev, sk); 9012 while (lower) { 9013 dev = lower; 9014 lower = netdev_sk_get_lower_dev(dev, sk); 9015 } 9016 9017 return dev; 9018 } 9019 EXPORT_SYMBOL(netdev_sk_get_lowest_dev); 9020 9021 static void netdev_adjacent_add_links(struct net_device *dev) 9022 { 9023 struct netdev_adjacent *iter; 9024 9025 struct net *net = dev_net(dev); 9026 9027 list_for_each_entry(iter, &dev->adj_list.upper, list) { 9028 if (!net_eq(net, dev_net(iter->dev))) 9029 continue; 9030 netdev_adjacent_sysfs_add(iter->dev, dev, 9031 &iter->dev->adj_list.lower); 9032 netdev_adjacent_sysfs_add(dev, iter->dev, 9033 &dev->adj_list.upper); 9034 } 9035 9036 list_for_each_entry(iter, &dev->adj_list.lower, list) { 9037 if (!net_eq(net, dev_net(iter->dev))) 9038 continue; 9039 netdev_adjacent_sysfs_add(iter->dev, dev, 9040 &iter->dev->adj_list.upper); 9041 netdev_adjacent_sysfs_add(dev, iter->dev, 9042 &dev->adj_list.lower); 9043 } 9044 } 9045 9046 static void netdev_adjacent_del_links(struct net_device *dev) 9047 { 9048 struct netdev_adjacent *iter; 9049 9050 struct net *net = dev_net(dev); 9051 9052 list_for_each_entry(iter, &dev->adj_list.upper, list) { 9053 if (!net_eq(net, dev_net(iter->dev))) 9054 continue; 9055 netdev_adjacent_sysfs_del(iter->dev, dev->name, 9056 &iter->dev->adj_list.lower); 9057 netdev_adjacent_sysfs_del(dev, iter->dev->name, 9058 &dev->adj_list.upper); 9059 } 9060 9061 list_for_each_entry(iter, &dev->adj_list.lower, list) { 9062 if (!net_eq(net, dev_net(iter->dev))) 9063 continue; 9064 netdev_adjacent_sysfs_del(iter->dev, dev->name, 9065 &iter->dev->adj_list.upper); 9066 netdev_adjacent_sysfs_del(dev, iter->dev->name, 9067 &dev->adj_list.lower); 9068 } 9069 } 9070 9071 void netdev_adjacent_rename_links(struct net_device *dev, char *oldname) 9072 { 9073 struct netdev_adjacent *iter; 9074 9075 struct net *net = dev_net(dev); 9076 9077 list_for_each_entry(iter, &dev->adj_list.upper, list) { 9078 if (!net_eq(net, dev_net(iter->dev))) 9079 continue; 9080 netdev_adjacent_sysfs_del(iter->dev, oldname, 9081 &iter->dev->adj_list.lower); 9082 netdev_adjacent_sysfs_add(iter->dev, dev, 9083 &iter->dev->adj_list.lower); 9084 } 9085 9086 list_for_each_entry(iter, &dev->adj_list.lower, list) { 9087 if (!net_eq(net, dev_net(iter->dev))) 9088 continue; 9089 netdev_adjacent_sysfs_del(iter->dev, oldname, 9090 &iter->dev->adj_list.upper); 9091 netdev_adjacent_sysfs_add(iter->dev, dev, 9092 &iter->dev->adj_list.upper); 9093 } 9094 } 9095 9096 void *netdev_lower_dev_get_private(struct net_device *dev, 9097 struct net_device *lower_dev) 9098 { 9099 struct netdev_adjacent *lower; 9100 9101 if (!lower_dev) 9102 return NULL; 9103 lower = __netdev_find_adj(lower_dev, &dev->adj_list.lower); 9104 if (!lower) 9105 return NULL; 9106 9107 return lower->private; 9108 } 9109 EXPORT_SYMBOL(netdev_lower_dev_get_private); 9110 9111 9112 /** 9113 * netdev_lower_state_changed - Dispatch event about lower device state change 9114 * @lower_dev: device 9115 * @lower_state_info: state to dispatch 9116 * 9117 * Send NETDEV_CHANGELOWERSTATE to netdev notifiers with info. 9118 * The caller must hold the RTNL lock. 9119 */ 9120 void netdev_lower_state_changed(struct net_device *lower_dev, 9121 void *lower_state_info) 9122 { 9123 struct netdev_notifier_changelowerstate_info changelowerstate_info = { 9124 .info.dev = lower_dev, 9125 }; 9126 9127 ASSERT_RTNL(); 9128 changelowerstate_info.lower_state_info = lower_state_info; 9129 call_netdevice_notifiers_info(NETDEV_CHANGELOWERSTATE, 9130 &changelowerstate_info.info); 9131 } 9132 EXPORT_SYMBOL(netdev_lower_state_changed); 9133 9134 static void dev_change_rx_flags(struct net_device *dev, int flags) 9135 { 9136 const struct net_device_ops *ops = dev->netdev_ops; 9137 9138 if (ops->ndo_change_rx_flags) 9139 ops->ndo_change_rx_flags(dev, flags); 9140 } 9141 9142 static int __dev_set_promiscuity(struct net_device *dev, int inc, bool notify) 9143 { 9144 unsigned int old_flags = dev->flags; 9145 unsigned int promiscuity, flags; 9146 kuid_t uid; 9147 kgid_t gid; 9148 9149 ASSERT_RTNL(); 9150 9151 promiscuity = dev->promiscuity + inc; 9152 if (promiscuity == 0) { 9153 /* 9154 * Avoid overflow. 9155 * If inc causes overflow, untouch promisc and return error. 9156 */ 9157 if (unlikely(inc > 0)) { 9158 netdev_warn(dev, "promiscuity touches roof, set promiscuity failed. promiscuity feature of device might be broken.\n"); 9159 return -EOVERFLOW; 9160 } 9161 flags = old_flags & ~IFF_PROMISC; 9162 } else { 9163 flags = old_flags | IFF_PROMISC; 9164 } 9165 WRITE_ONCE(dev->promiscuity, promiscuity); 9166 if (flags != old_flags) { 9167 WRITE_ONCE(dev->flags, flags); 9168 netdev_info(dev, "%s promiscuous mode\n", 9169 dev->flags & IFF_PROMISC ? "entered" : "left"); 9170 if (audit_enabled) { 9171 current_uid_gid(&uid, &gid); 9172 audit_log(audit_context(), GFP_ATOMIC, 9173 AUDIT_ANOM_PROMISCUOUS, 9174 "dev=%s prom=%d old_prom=%d auid=%u uid=%u gid=%u ses=%u", 9175 dev->name, (dev->flags & IFF_PROMISC), 9176 (old_flags & IFF_PROMISC), 9177 from_kuid(&init_user_ns, audit_get_loginuid(current)), 9178 from_kuid(&init_user_ns, uid), 9179 from_kgid(&init_user_ns, gid), 9180 audit_get_sessionid(current)); 9181 } 9182 9183 dev_change_rx_flags(dev, IFF_PROMISC); 9184 } 9185 if (notify) 9186 __dev_notify_flags(dev, old_flags, IFF_PROMISC, 0, NULL); 9187 return 0; 9188 } 9189 9190 /** 9191 * dev_set_promiscuity - update promiscuity count on a device 9192 * @dev: device 9193 * @inc: modifier 9194 * 9195 * Add or remove promiscuity from a device. While the count in the device 9196 * remains above zero the interface remains promiscuous. Once it hits zero 9197 * the device reverts back to normal filtering operation. A negative inc 9198 * value is used to drop promiscuity on the device. 9199 * Return 0 if successful or a negative errno code on error. 9200 */ 9201 int dev_set_promiscuity(struct net_device *dev, int inc) 9202 { 9203 unsigned int old_flags = dev->flags; 9204 int err; 9205 9206 err = __dev_set_promiscuity(dev, inc, true); 9207 if (err < 0) 9208 return err; 9209 if (dev->flags != old_flags) 9210 dev_set_rx_mode(dev); 9211 return err; 9212 } 9213 EXPORT_SYMBOL(dev_set_promiscuity); 9214 9215 int netif_set_allmulti(struct net_device *dev, int inc, bool notify) 9216 { 9217 unsigned int old_flags = dev->flags, old_gflags = dev->gflags; 9218 unsigned int allmulti, flags; 9219 9220 ASSERT_RTNL(); 9221 9222 allmulti = dev->allmulti + inc; 9223 if (allmulti == 0) { 9224 /* 9225 * Avoid overflow. 9226 * If inc causes overflow, untouch allmulti and return error. 9227 */ 9228 if (unlikely(inc > 0)) { 9229 netdev_warn(dev, "allmulti touches roof, set allmulti failed. allmulti feature of device might be broken.\n"); 9230 return -EOVERFLOW; 9231 } 9232 flags = old_flags & ~IFF_ALLMULTI; 9233 } else { 9234 flags = old_flags | IFF_ALLMULTI; 9235 } 9236 WRITE_ONCE(dev->allmulti, allmulti); 9237 if (flags != old_flags) { 9238 WRITE_ONCE(dev->flags, flags); 9239 netdev_info(dev, "%s allmulticast mode\n", 9240 dev->flags & IFF_ALLMULTI ? "entered" : "left"); 9241 dev_change_rx_flags(dev, IFF_ALLMULTI); 9242 dev_set_rx_mode(dev); 9243 if (notify) 9244 __dev_notify_flags(dev, old_flags, 9245 dev->gflags ^ old_gflags, 0, NULL); 9246 } 9247 return 0; 9248 } 9249 9250 /* 9251 * Upload unicast and multicast address lists to device and 9252 * configure RX filtering. When the device doesn't support unicast 9253 * filtering it is put in promiscuous mode while unicast addresses 9254 * are present. 9255 */ 9256 void __dev_set_rx_mode(struct net_device *dev) 9257 { 9258 const struct net_device_ops *ops = dev->netdev_ops; 9259 9260 /* dev_open will call this function so the list will stay sane. */ 9261 if (!(dev->flags&IFF_UP)) 9262 return; 9263 9264 if (!netif_device_present(dev)) 9265 return; 9266 9267 if (!(dev->priv_flags & IFF_UNICAST_FLT)) { 9268 /* Unicast addresses changes may only happen under the rtnl, 9269 * therefore calling __dev_set_promiscuity here is safe. 9270 */ 9271 if (!netdev_uc_empty(dev) && !dev->uc_promisc) { 9272 __dev_set_promiscuity(dev, 1, false); 9273 dev->uc_promisc = true; 9274 } else if (netdev_uc_empty(dev) && dev->uc_promisc) { 9275 __dev_set_promiscuity(dev, -1, false); 9276 dev->uc_promisc = false; 9277 } 9278 } 9279 9280 if (ops->ndo_set_rx_mode) 9281 ops->ndo_set_rx_mode(dev); 9282 } 9283 9284 void dev_set_rx_mode(struct net_device *dev) 9285 { 9286 netif_addr_lock_bh(dev); 9287 __dev_set_rx_mode(dev); 9288 netif_addr_unlock_bh(dev); 9289 } 9290 9291 /** 9292 * dev_get_flags - get flags reported to userspace 9293 * @dev: device 9294 * 9295 * Get the combination of flag bits exported through APIs to userspace. 9296 */ 9297 unsigned int dev_get_flags(const struct net_device *dev) 9298 { 9299 unsigned int flags; 9300 9301 flags = (READ_ONCE(dev->flags) & ~(IFF_PROMISC | 9302 IFF_ALLMULTI | 9303 IFF_RUNNING | 9304 IFF_LOWER_UP | 9305 IFF_DORMANT)) | 9306 (READ_ONCE(dev->gflags) & (IFF_PROMISC | 9307 IFF_ALLMULTI)); 9308 9309 if (netif_running(dev)) { 9310 if (netif_oper_up(dev)) 9311 flags |= IFF_RUNNING; 9312 if (netif_carrier_ok(dev)) 9313 flags |= IFF_LOWER_UP; 9314 if (netif_dormant(dev)) 9315 flags |= IFF_DORMANT; 9316 } 9317 9318 return flags; 9319 } 9320 EXPORT_SYMBOL(dev_get_flags); 9321 9322 int __dev_change_flags(struct net_device *dev, unsigned int flags, 9323 struct netlink_ext_ack *extack) 9324 { 9325 unsigned int old_flags = dev->flags; 9326 int ret; 9327 9328 ASSERT_RTNL(); 9329 9330 /* 9331 * Set the flags on our device. 9332 */ 9333 9334 dev->flags = (flags & (IFF_DEBUG | IFF_NOTRAILERS | IFF_NOARP | 9335 IFF_DYNAMIC | IFF_MULTICAST | IFF_PORTSEL | 9336 IFF_AUTOMEDIA)) | 9337 (dev->flags & (IFF_UP | IFF_VOLATILE | IFF_PROMISC | 9338 IFF_ALLMULTI)); 9339 9340 /* 9341 * Load in the correct multicast list now the flags have changed. 9342 */ 9343 9344 if ((old_flags ^ flags) & IFF_MULTICAST) 9345 dev_change_rx_flags(dev, IFF_MULTICAST); 9346 9347 dev_set_rx_mode(dev); 9348 9349 /* 9350 * Have we downed the interface. We handle IFF_UP ourselves 9351 * according to user attempts to set it, rather than blindly 9352 * setting it. 9353 */ 9354 9355 ret = 0; 9356 if ((old_flags ^ flags) & IFF_UP) { 9357 if (old_flags & IFF_UP) 9358 __dev_close(dev); 9359 else 9360 ret = __dev_open(dev, extack); 9361 } 9362 9363 if ((flags ^ dev->gflags) & IFF_PROMISC) { 9364 int inc = (flags & IFF_PROMISC) ? 1 : -1; 9365 old_flags = dev->flags; 9366 9367 dev->gflags ^= IFF_PROMISC; 9368 9369 if (__dev_set_promiscuity(dev, inc, false) >= 0) 9370 if (dev->flags != old_flags) 9371 dev_set_rx_mode(dev); 9372 } 9373 9374 /* NOTE: order of synchronization of IFF_PROMISC and IFF_ALLMULTI 9375 * is important. Some (broken) drivers set IFF_PROMISC, when 9376 * IFF_ALLMULTI is requested not asking us and not reporting. 9377 */ 9378 if ((flags ^ dev->gflags) & IFF_ALLMULTI) { 9379 int inc = (flags & IFF_ALLMULTI) ? 1 : -1; 9380 9381 dev->gflags ^= IFF_ALLMULTI; 9382 netif_set_allmulti(dev, inc, false); 9383 } 9384 9385 return ret; 9386 } 9387 9388 void __dev_notify_flags(struct net_device *dev, unsigned int old_flags, 9389 unsigned int gchanges, u32 portid, 9390 const struct nlmsghdr *nlh) 9391 { 9392 unsigned int changes = dev->flags ^ old_flags; 9393 9394 if (gchanges) 9395 rtmsg_ifinfo(RTM_NEWLINK, dev, gchanges, GFP_ATOMIC, portid, nlh); 9396 9397 if (changes & IFF_UP) { 9398 if (dev->flags & IFF_UP) 9399 call_netdevice_notifiers(NETDEV_UP, dev); 9400 else 9401 call_netdevice_notifiers(NETDEV_DOWN, dev); 9402 } 9403 9404 if (dev->flags & IFF_UP && 9405 (changes & ~(IFF_UP | IFF_PROMISC | IFF_ALLMULTI | IFF_VOLATILE))) { 9406 struct netdev_notifier_change_info change_info = { 9407 .info = { 9408 .dev = dev, 9409 }, 9410 .flags_changed = changes, 9411 }; 9412 9413 call_netdevice_notifiers_info(NETDEV_CHANGE, &change_info.info); 9414 } 9415 } 9416 9417 int netif_change_flags(struct net_device *dev, unsigned int flags, 9418 struct netlink_ext_ack *extack) 9419 { 9420 int ret; 9421 unsigned int changes, old_flags = dev->flags, old_gflags = dev->gflags; 9422 9423 ret = __dev_change_flags(dev, flags, extack); 9424 if (ret < 0) 9425 return ret; 9426 9427 changes = (old_flags ^ dev->flags) | (old_gflags ^ dev->gflags); 9428 __dev_notify_flags(dev, old_flags, changes, 0, NULL); 9429 return ret; 9430 } 9431 9432 int __dev_set_mtu(struct net_device *dev, int new_mtu) 9433 { 9434 const struct net_device_ops *ops = dev->netdev_ops; 9435 9436 if (ops->ndo_change_mtu) 9437 return ops->ndo_change_mtu(dev, new_mtu); 9438 9439 /* Pairs with all the lockless reads of dev->mtu in the stack */ 9440 WRITE_ONCE(dev->mtu, new_mtu); 9441 return 0; 9442 } 9443 EXPORT_SYMBOL(__dev_set_mtu); 9444 9445 int dev_validate_mtu(struct net_device *dev, int new_mtu, 9446 struct netlink_ext_ack *extack) 9447 { 9448 /* MTU must be positive, and in range */ 9449 if (new_mtu < 0 || new_mtu < dev->min_mtu) { 9450 NL_SET_ERR_MSG(extack, "mtu less than device minimum"); 9451 return -EINVAL; 9452 } 9453 9454 if (dev->max_mtu > 0 && new_mtu > dev->max_mtu) { 9455 NL_SET_ERR_MSG(extack, "mtu greater than device maximum"); 9456 return -EINVAL; 9457 } 9458 return 0; 9459 } 9460 9461 /** 9462 * netif_set_mtu_ext - Change maximum transfer unit 9463 * @dev: device 9464 * @new_mtu: new transfer unit 9465 * @extack: netlink extended ack 9466 * 9467 * Change the maximum transfer size of the network device. 9468 */ 9469 int netif_set_mtu_ext(struct net_device *dev, int new_mtu, 9470 struct netlink_ext_ack *extack) 9471 { 9472 int err, orig_mtu; 9473 9474 if (new_mtu == dev->mtu) 9475 return 0; 9476 9477 err = dev_validate_mtu(dev, new_mtu, extack); 9478 if (err) 9479 return err; 9480 9481 if (!netif_device_present(dev)) 9482 return -ENODEV; 9483 9484 err = call_netdevice_notifiers(NETDEV_PRECHANGEMTU, dev); 9485 err = notifier_to_errno(err); 9486 if (err) 9487 return err; 9488 9489 orig_mtu = dev->mtu; 9490 err = __dev_set_mtu(dev, new_mtu); 9491 9492 if (!err) { 9493 err = call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev, 9494 orig_mtu); 9495 err = notifier_to_errno(err); 9496 if (err) { 9497 /* setting mtu back and notifying everyone again, 9498 * so that they have a chance to revert changes. 9499 */ 9500 __dev_set_mtu(dev, orig_mtu); 9501 call_netdevice_notifiers_mtu(NETDEV_CHANGEMTU, dev, 9502 new_mtu); 9503 } 9504 } 9505 return err; 9506 } 9507 9508 int netif_set_mtu(struct net_device *dev, int new_mtu) 9509 { 9510 struct netlink_ext_ack extack; 9511 int err; 9512 9513 memset(&extack, 0, sizeof(extack)); 9514 err = netif_set_mtu_ext(dev, new_mtu, &extack); 9515 if (err && extack._msg) 9516 net_err_ratelimited("%s: %s\n", dev->name, extack._msg); 9517 return err; 9518 } 9519 EXPORT_SYMBOL(netif_set_mtu); 9520 9521 int netif_change_tx_queue_len(struct net_device *dev, unsigned long new_len) 9522 { 9523 unsigned int orig_len = dev->tx_queue_len; 9524 int res; 9525 9526 if (new_len != (unsigned int)new_len) 9527 return -ERANGE; 9528 9529 if (new_len != orig_len) { 9530 WRITE_ONCE(dev->tx_queue_len, new_len); 9531 res = call_netdevice_notifiers(NETDEV_CHANGE_TX_QUEUE_LEN, dev); 9532 res = notifier_to_errno(res); 9533 if (res) 9534 goto err_rollback; 9535 res = dev_qdisc_change_tx_queue_len(dev); 9536 if (res) 9537 goto err_rollback; 9538 } 9539 9540 return 0; 9541 9542 err_rollback: 9543 netdev_err(dev, "refused to change device tx_queue_len\n"); 9544 WRITE_ONCE(dev->tx_queue_len, orig_len); 9545 return res; 9546 } 9547 9548 void netif_set_group(struct net_device *dev, int new_group) 9549 { 9550 dev->group = new_group; 9551 } 9552 9553 /** 9554 * dev_pre_changeaddr_notify - Call NETDEV_PRE_CHANGEADDR. 9555 * @dev: device 9556 * @addr: new address 9557 * @extack: netlink extended ack 9558 */ 9559 int dev_pre_changeaddr_notify(struct net_device *dev, const char *addr, 9560 struct netlink_ext_ack *extack) 9561 { 9562 struct netdev_notifier_pre_changeaddr_info info = { 9563 .info.dev = dev, 9564 .info.extack = extack, 9565 .dev_addr = addr, 9566 }; 9567 int rc; 9568 9569 rc = call_netdevice_notifiers_info(NETDEV_PRE_CHANGEADDR, &info.info); 9570 return notifier_to_errno(rc); 9571 } 9572 EXPORT_SYMBOL(dev_pre_changeaddr_notify); 9573 9574 int netif_set_mac_address(struct net_device *dev, struct sockaddr *sa, 9575 struct netlink_ext_ack *extack) 9576 { 9577 const struct net_device_ops *ops = dev->netdev_ops; 9578 int err; 9579 9580 if (!ops->ndo_set_mac_address) 9581 return -EOPNOTSUPP; 9582 if (sa->sa_family != dev->type) 9583 return -EINVAL; 9584 if (!netif_device_present(dev)) 9585 return -ENODEV; 9586 err = dev_pre_changeaddr_notify(dev, sa->sa_data, extack); 9587 if (err) 9588 return err; 9589 if (memcmp(dev->dev_addr, sa->sa_data, dev->addr_len)) { 9590 err = ops->ndo_set_mac_address(dev, sa); 9591 if (err) 9592 return err; 9593 } 9594 dev->addr_assign_type = NET_ADDR_SET; 9595 call_netdevice_notifiers(NETDEV_CHANGEADDR, dev); 9596 add_device_randomness(dev->dev_addr, dev->addr_len); 9597 return 0; 9598 } 9599 9600 int dev_get_mac_address(struct sockaddr *sa, struct net *net, char *dev_name) 9601 { 9602 size_t size = sizeof(sa->sa_data_min); 9603 struct net_device *dev; 9604 9605 dev = netdev_get_by_name_lock(net, dev_name); 9606 if (!dev) 9607 return -ENODEV; 9608 9609 if (!dev->addr_len) 9610 memset(sa->sa_data, 0, size); 9611 else 9612 memcpy(sa->sa_data, dev->dev_addr, 9613 min_t(size_t, size, dev->addr_len)); 9614 sa->sa_family = dev->type; 9615 netdev_unlock(dev); 9616 9617 return 0; 9618 } 9619 EXPORT_SYMBOL(dev_get_mac_address); 9620 9621 int netif_change_carrier(struct net_device *dev, bool new_carrier) 9622 { 9623 const struct net_device_ops *ops = dev->netdev_ops; 9624 9625 if (!ops->ndo_change_carrier) 9626 return -EOPNOTSUPP; 9627 if (!netif_device_present(dev)) 9628 return -ENODEV; 9629 return ops->ndo_change_carrier(dev, new_carrier); 9630 } 9631 9632 /** 9633 * dev_get_phys_port_id - Get device physical port ID 9634 * @dev: device 9635 * @ppid: port ID 9636 * 9637 * Get device physical port ID 9638 */ 9639 int dev_get_phys_port_id(struct net_device *dev, 9640 struct netdev_phys_item_id *ppid) 9641 { 9642 const struct net_device_ops *ops = dev->netdev_ops; 9643 9644 if (!ops->ndo_get_phys_port_id) 9645 return -EOPNOTSUPP; 9646 return ops->ndo_get_phys_port_id(dev, ppid); 9647 } 9648 9649 /** 9650 * dev_get_phys_port_name - Get device physical port name 9651 * @dev: device 9652 * @name: port name 9653 * @len: limit of bytes to copy to name 9654 * 9655 * Get device physical port name 9656 */ 9657 int dev_get_phys_port_name(struct net_device *dev, 9658 char *name, size_t len) 9659 { 9660 const struct net_device_ops *ops = dev->netdev_ops; 9661 int err; 9662 9663 if (ops->ndo_get_phys_port_name) { 9664 err = ops->ndo_get_phys_port_name(dev, name, len); 9665 if (err != -EOPNOTSUPP) 9666 return err; 9667 } 9668 return devlink_compat_phys_port_name_get(dev, name, len); 9669 } 9670 9671 /** 9672 * dev_get_port_parent_id - Get the device's port parent identifier 9673 * @dev: network device 9674 * @ppid: pointer to a storage for the port's parent identifier 9675 * @recurse: allow/disallow recursion to lower devices 9676 * 9677 * Get the devices's port parent identifier 9678 */ 9679 int dev_get_port_parent_id(struct net_device *dev, 9680 struct netdev_phys_item_id *ppid, 9681 bool recurse) 9682 { 9683 const struct net_device_ops *ops = dev->netdev_ops; 9684 struct netdev_phys_item_id first = { }; 9685 struct net_device *lower_dev; 9686 struct list_head *iter; 9687 int err; 9688 9689 if (ops->ndo_get_port_parent_id) { 9690 err = ops->ndo_get_port_parent_id(dev, ppid); 9691 if (err != -EOPNOTSUPP) 9692 return err; 9693 } 9694 9695 err = devlink_compat_switch_id_get(dev, ppid); 9696 if (!recurse || err != -EOPNOTSUPP) 9697 return err; 9698 9699 netdev_for_each_lower_dev(dev, lower_dev, iter) { 9700 err = dev_get_port_parent_id(lower_dev, ppid, true); 9701 if (err) 9702 break; 9703 if (!first.id_len) 9704 first = *ppid; 9705 else if (memcmp(&first, ppid, sizeof(*ppid))) 9706 return -EOPNOTSUPP; 9707 } 9708 9709 return err; 9710 } 9711 EXPORT_SYMBOL(dev_get_port_parent_id); 9712 9713 /** 9714 * netdev_port_same_parent_id - Indicate if two network devices have 9715 * the same port parent identifier 9716 * @a: first network device 9717 * @b: second network device 9718 */ 9719 bool netdev_port_same_parent_id(struct net_device *a, struct net_device *b) 9720 { 9721 struct netdev_phys_item_id a_id = { }; 9722 struct netdev_phys_item_id b_id = { }; 9723 9724 if (dev_get_port_parent_id(a, &a_id, true) || 9725 dev_get_port_parent_id(b, &b_id, true)) 9726 return false; 9727 9728 return netdev_phys_item_id_same(&a_id, &b_id); 9729 } 9730 EXPORT_SYMBOL(netdev_port_same_parent_id); 9731 9732 int netif_change_proto_down(struct net_device *dev, bool proto_down) 9733 { 9734 if (!dev->change_proto_down) 9735 return -EOPNOTSUPP; 9736 if (!netif_device_present(dev)) 9737 return -ENODEV; 9738 if (proto_down) 9739 netif_carrier_off(dev); 9740 else 9741 netif_carrier_on(dev); 9742 WRITE_ONCE(dev->proto_down, proto_down); 9743 return 0; 9744 } 9745 9746 /** 9747 * netdev_change_proto_down_reason_locked - proto down reason 9748 * 9749 * @dev: device 9750 * @mask: proto down mask 9751 * @value: proto down value 9752 */ 9753 void netdev_change_proto_down_reason_locked(struct net_device *dev, 9754 unsigned long mask, u32 value) 9755 { 9756 u32 proto_down_reason; 9757 int b; 9758 9759 if (!mask) { 9760 proto_down_reason = value; 9761 } else { 9762 proto_down_reason = dev->proto_down_reason; 9763 for_each_set_bit(b, &mask, 32) { 9764 if (value & (1 << b)) 9765 proto_down_reason |= BIT(b); 9766 else 9767 proto_down_reason &= ~BIT(b); 9768 } 9769 } 9770 WRITE_ONCE(dev->proto_down_reason, proto_down_reason); 9771 } 9772 9773 struct bpf_xdp_link { 9774 struct bpf_link link; 9775 struct net_device *dev; /* protected by rtnl_lock, no refcnt held */ 9776 int flags; 9777 }; 9778 9779 static enum bpf_xdp_mode dev_xdp_mode(struct net_device *dev, u32 flags) 9780 { 9781 if (flags & XDP_FLAGS_HW_MODE) 9782 return XDP_MODE_HW; 9783 if (flags & XDP_FLAGS_DRV_MODE) 9784 return XDP_MODE_DRV; 9785 if (flags & XDP_FLAGS_SKB_MODE) 9786 return XDP_MODE_SKB; 9787 return dev->netdev_ops->ndo_bpf ? XDP_MODE_DRV : XDP_MODE_SKB; 9788 } 9789 9790 static bpf_op_t dev_xdp_bpf_op(struct net_device *dev, enum bpf_xdp_mode mode) 9791 { 9792 switch (mode) { 9793 case XDP_MODE_SKB: 9794 return generic_xdp_install; 9795 case XDP_MODE_DRV: 9796 case XDP_MODE_HW: 9797 return dev->netdev_ops->ndo_bpf; 9798 default: 9799 return NULL; 9800 } 9801 } 9802 9803 static struct bpf_xdp_link *dev_xdp_link(struct net_device *dev, 9804 enum bpf_xdp_mode mode) 9805 { 9806 return dev->xdp_state[mode].link; 9807 } 9808 9809 static struct bpf_prog *dev_xdp_prog(struct net_device *dev, 9810 enum bpf_xdp_mode mode) 9811 { 9812 struct bpf_xdp_link *link = dev_xdp_link(dev, mode); 9813 9814 if (link) 9815 return link->link.prog; 9816 return dev->xdp_state[mode].prog; 9817 } 9818 9819 u8 dev_xdp_prog_count(struct net_device *dev) 9820 { 9821 u8 count = 0; 9822 int i; 9823 9824 for (i = 0; i < __MAX_XDP_MODE; i++) 9825 if (dev->xdp_state[i].prog || dev->xdp_state[i].link) 9826 count++; 9827 return count; 9828 } 9829 EXPORT_SYMBOL_GPL(dev_xdp_prog_count); 9830 9831 u8 dev_xdp_sb_prog_count(struct net_device *dev) 9832 { 9833 u8 count = 0; 9834 int i; 9835 9836 for (i = 0; i < __MAX_XDP_MODE; i++) 9837 if (dev->xdp_state[i].prog && 9838 !dev->xdp_state[i].prog->aux->xdp_has_frags) 9839 count++; 9840 return count; 9841 } 9842 9843 int netif_xdp_propagate(struct net_device *dev, struct netdev_bpf *bpf) 9844 { 9845 if (!dev->netdev_ops->ndo_bpf) 9846 return -EOPNOTSUPP; 9847 9848 if (dev->cfg->hds_config == ETHTOOL_TCP_DATA_SPLIT_ENABLED && 9849 bpf->command == XDP_SETUP_PROG && 9850 bpf->prog && !bpf->prog->aux->xdp_has_frags) { 9851 NL_SET_ERR_MSG(bpf->extack, 9852 "unable to propagate XDP to device using tcp-data-split"); 9853 return -EBUSY; 9854 } 9855 9856 if (dev_get_min_mp_channel_count(dev)) { 9857 NL_SET_ERR_MSG(bpf->extack, "unable to propagate XDP to device using memory provider"); 9858 return -EBUSY; 9859 } 9860 9861 return dev->netdev_ops->ndo_bpf(dev, bpf); 9862 } 9863 9864 u32 dev_xdp_prog_id(struct net_device *dev, enum bpf_xdp_mode mode) 9865 { 9866 struct bpf_prog *prog = dev_xdp_prog(dev, mode); 9867 9868 return prog ? prog->aux->id : 0; 9869 } 9870 9871 static void dev_xdp_set_link(struct net_device *dev, enum bpf_xdp_mode mode, 9872 struct bpf_xdp_link *link) 9873 { 9874 dev->xdp_state[mode].link = link; 9875 dev->xdp_state[mode].prog = NULL; 9876 } 9877 9878 static void dev_xdp_set_prog(struct net_device *dev, enum bpf_xdp_mode mode, 9879 struct bpf_prog *prog) 9880 { 9881 dev->xdp_state[mode].link = NULL; 9882 dev->xdp_state[mode].prog = prog; 9883 } 9884 9885 static int dev_xdp_install(struct net_device *dev, enum bpf_xdp_mode mode, 9886 bpf_op_t bpf_op, struct netlink_ext_ack *extack, 9887 u32 flags, struct bpf_prog *prog) 9888 { 9889 struct netdev_bpf xdp; 9890 int err; 9891 9892 netdev_ops_assert_locked(dev); 9893 9894 if (dev->cfg->hds_config == ETHTOOL_TCP_DATA_SPLIT_ENABLED && 9895 prog && !prog->aux->xdp_has_frags) { 9896 NL_SET_ERR_MSG(extack, "unable to install XDP to device using tcp-data-split"); 9897 return -EBUSY; 9898 } 9899 9900 if (dev_get_min_mp_channel_count(dev)) { 9901 NL_SET_ERR_MSG(extack, "unable to install XDP to device using memory provider"); 9902 return -EBUSY; 9903 } 9904 9905 memset(&xdp, 0, sizeof(xdp)); 9906 xdp.command = mode == XDP_MODE_HW ? XDP_SETUP_PROG_HW : XDP_SETUP_PROG; 9907 xdp.extack = extack; 9908 xdp.flags = flags; 9909 xdp.prog = prog; 9910 9911 /* Drivers assume refcnt is already incremented (i.e, prog pointer is 9912 * "moved" into driver), so they don't increment it on their own, but 9913 * they do decrement refcnt when program is detached or replaced. 9914 * Given net_device also owns link/prog, we need to bump refcnt here 9915 * to prevent drivers from underflowing it. 9916 */ 9917 if (prog) 9918 bpf_prog_inc(prog); 9919 err = bpf_op(dev, &xdp); 9920 if (err) { 9921 if (prog) 9922 bpf_prog_put(prog); 9923 return err; 9924 } 9925 9926 if (mode != XDP_MODE_HW) 9927 bpf_prog_change_xdp(dev_xdp_prog(dev, mode), prog); 9928 9929 return 0; 9930 } 9931 9932 static void dev_xdp_uninstall(struct net_device *dev) 9933 { 9934 struct bpf_xdp_link *link; 9935 struct bpf_prog *prog; 9936 enum bpf_xdp_mode mode; 9937 bpf_op_t bpf_op; 9938 9939 ASSERT_RTNL(); 9940 9941 for (mode = XDP_MODE_SKB; mode < __MAX_XDP_MODE; mode++) { 9942 prog = dev_xdp_prog(dev, mode); 9943 if (!prog) 9944 continue; 9945 9946 bpf_op = dev_xdp_bpf_op(dev, mode); 9947 if (!bpf_op) 9948 continue; 9949 9950 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL)); 9951 9952 /* auto-detach link from net device */ 9953 link = dev_xdp_link(dev, mode); 9954 if (link) 9955 link->dev = NULL; 9956 else 9957 bpf_prog_put(prog); 9958 9959 dev_xdp_set_link(dev, mode, NULL); 9960 } 9961 } 9962 9963 static int dev_xdp_attach(struct net_device *dev, struct netlink_ext_ack *extack, 9964 struct bpf_xdp_link *link, struct bpf_prog *new_prog, 9965 struct bpf_prog *old_prog, u32 flags) 9966 { 9967 unsigned int num_modes = hweight32(flags & XDP_FLAGS_MODES); 9968 struct bpf_prog *cur_prog; 9969 struct net_device *upper; 9970 struct list_head *iter; 9971 enum bpf_xdp_mode mode; 9972 bpf_op_t bpf_op; 9973 int err; 9974 9975 ASSERT_RTNL(); 9976 9977 /* either link or prog attachment, never both */ 9978 if (link && (new_prog || old_prog)) 9979 return -EINVAL; 9980 /* link supports only XDP mode flags */ 9981 if (link && (flags & ~XDP_FLAGS_MODES)) { 9982 NL_SET_ERR_MSG(extack, "Invalid XDP flags for BPF link attachment"); 9983 return -EINVAL; 9984 } 9985 /* just one XDP mode bit should be set, zero defaults to drv/skb mode */ 9986 if (num_modes > 1) { 9987 NL_SET_ERR_MSG(extack, "Only one XDP mode flag can be set"); 9988 return -EINVAL; 9989 } 9990 /* avoid ambiguity if offload + drv/skb mode progs are both loaded */ 9991 if (!num_modes && dev_xdp_prog_count(dev) > 1) { 9992 NL_SET_ERR_MSG(extack, 9993 "More than one program loaded, unset mode is ambiguous"); 9994 return -EINVAL; 9995 } 9996 /* old_prog != NULL implies XDP_FLAGS_REPLACE is set */ 9997 if (old_prog && !(flags & XDP_FLAGS_REPLACE)) { 9998 NL_SET_ERR_MSG(extack, "XDP_FLAGS_REPLACE is not specified"); 9999 return -EINVAL; 10000 } 10001 10002 mode = dev_xdp_mode(dev, flags); 10003 /* can't replace attached link */ 10004 if (dev_xdp_link(dev, mode)) { 10005 NL_SET_ERR_MSG(extack, "Can't replace active BPF XDP link"); 10006 return -EBUSY; 10007 } 10008 10009 /* don't allow if an upper device already has a program */ 10010 netdev_for_each_upper_dev_rcu(dev, upper, iter) { 10011 if (dev_xdp_prog_count(upper) > 0) { 10012 NL_SET_ERR_MSG(extack, "Cannot attach when an upper device already has a program"); 10013 return -EEXIST; 10014 } 10015 } 10016 10017 cur_prog = dev_xdp_prog(dev, mode); 10018 /* can't replace attached prog with link */ 10019 if (link && cur_prog) { 10020 NL_SET_ERR_MSG(extack, "Can't replace active XDP program with BPF link"); 10021 return -EBUSY; 10022 } 10023 if ((flags & XDP_FLAGS_REPLACE) && cur_prog != old_prog) { 10024 NL_SET_ERR_MSG(extack, "Active program does not match expected"); 10025 return -EEXIST; 10026 } 10027 10028 /* put effective new program into new_prog */ 10029 if (link) 10030 new_prog = link->link.prog; 10031 10032 if (new_prog) { 10033 bool offload = mode == XDP_MODE_HW; 10034 enum bpf_xdp_mode other_mode = mode == XDP_MODE_SKB 10035 ? XDP_MODE_DRV : XDP_MODE_SKB; 10036 10037 if ((flags & XDP_FLAGS_UPDATE_IF_NOEXIST) && cur_prog) { 10038 NL_SET_ERR_MSG(extack, "XDP program already attached"); 10039 return -EBUSY; 10040 } 10041 if (!offload && dev_xdp_prog(dev, other_mode)) { 10042 NL_SET_ERR_MSG(extack, "Native and generic XDP can't be active at the same time"); 10043 return -EEXIST; 10044 } 10045 if (!offload && bpf_prog_is_offloaded(new_prog->aux)) { 10046 NL_SET_ERR_MSG(extack, "Using offloaded program without HW_MODE flag is not supported"); 10047 return -EINVAL; 10048 } 10049 if (bpf_prog_is_dev_bound(new_prog->aux) && !bpf_offload_dev_match(new_prog, dev)) { 10050 NL_SET_ERR_MSG(extack, "Program bound to different device"); 10051 return -EINVAL; 10052 } 10053 if (bpf_prog_is_dev_bound(new_prog->aux) && mode == XDP_MODE_SKB) { 10054 NL_SET_ERR_MSG(extack, "Can't attach device-bound programs in generic mode"); 10055 return -EINVAL; 10056 } 10057 if (new_prog->expected_attach_type == BPF_XDP_DEVMAP) { 10058 NL_SET_ERR_MSG(extack, "BPF_XDP_DEVMAP programs can not be attached to a device"); 10059 return -EINVAL; 10060 } 10061 if (new_prog->expected_attach_type == BPF_XDP_CPUMAP) { 10062 NL_SET_ERR_MSG(extack, "BPF_XDP_CPUMAP programs can not be attached to a device"); 10063 return -EINVAL; 10064 } 10065 } 10066 10067 /* don't call drivers if the effective program didn't change */ 10068 if (new_prog != cur_prog) { 10069 bpf_op = dev_xdp_bpf_op(dev, mode); 10070 if (!bpf_op) { 10071 NL_SET_ERR_MSG(extack, "Underlying driver does not support XDP in native mode"); 10072 return -EOPNOTSUPP; 10073 } 10074 10075 err = dev_xdp_install(dev, mode, bpf_op, extack, flags, new_prog); 10076 if (err) 10077 return err; 10078 } 10079 10080 if (link) 10081 dev_xdp_set_link(dev, mode, link); 10082 else 10083 dev_xdp_set_prog(dev, mode, new_prog); 10084 if (cur_prog) 10085 bpf_prog_put(cur_prog); 10086 10087 return 0; 10088 } 10089 10090 static int dev_xdp_attach_link(struct net_device *dev, 10091 struct netlink_ext_ack *extack, 10092 struct bpf_xdp_link *link) 10093 { 10094 return dev_xdp_attach(dev, extack, link, NULL, NULL, link->flags); 10095 } 10096 10097 static int dev_xdp_detach_link(struct net_device *dev, 10098 struct netlink_ext_ack *extack, 10099 struct bpf_xdp_link *link) 10100 { 10101 enum bpf_xdp_mode mode; 10102 bpf_op_t bpf_op; 10103 10104 ASSERT_RTNL(); 10105 10106 mode = dev_xdp_mode(dev, link->flags); 10107 if (dev_xdp_link(dev, mode) != link) 10108 return -EINVAL; 10109 10110 bpf_op = dev_xdp_bpf_op(dev, mode); 10111 WARN_ON(dev_xdp_install(dev, mode, bpf_op, NULL, 0, NULL)); 10112 dev_xdp_set_link(dev, mode, NULL); 10113 return 0; 10114 } 10115 10116 static void bpf_xdp_link_release(struct bpf_link *link) 10117 { 10118 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10119 10120 rtnl_lock(); 10121 10122 /* if racing with net_device's tear down, xdp_link->dev might be 10123 * already NULL, in which case link was already auto-detached 10124 */ 10125 if (xdp_link->dev) { 10126 netdev_lock_ops(xdp_link->dev); 10127 WARN_ON(dev_xdp_detach_link(xdp_link->dev, NULL, xdp_link)); 10128 netdev_unlock_ops(xdp_link->dev); 10129 xdp_link->dev = NULL; 10130 } 10131 10132 rtnl_unlock(); 10133 } 10134 10135 static int bpf_xdp_link_detach(struct bpf_link *link) 10136 { 10137 bpf_xdp_link_release(link); 10138 return 0; 10139 } 10140 10141 static void bpf_xdp_link_dealloc(struct bpf_link *link) 10142 { 10143 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10144 10145 kfree(xdp_link); 10146 } 10147 10148 static void bpf_xdp_link_show_fdinfo(const struct bpf_link *link, 10149 struct seq_file *seq) 10150 { 10151 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10152 u32 ifindex = 0; 10153 10154 rtnl_lock(); 10155 if (xdp_link->dev) 10156 ifindex = xdp_link->dev->ifindex; 10157 rtnl_unlock(); 10158 10159 seq_printf(seq, "ifindex:\t%u\n", ifindex); 10160 } 10161 10162 static int bpf_xdp_link_fill_link_info(const struct bpf_link *link, 10163 struct bpf_link_info *info) 10164 { 10165 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10166 u32 ifindex = 0; 10167 10168 rtnl_lock(); 10169 if (xdp_link->dev) 10170 ifindex = xdp_link->dev->ifindex; 10171 rtnl_unlock(); 10172 10173 info->xdp.ifindex = ifindex; 10174 return 0; 10175 } 10176 10177 static int bpf_xdp_link_update(struct bpf_link *link, struct bpf_prog *new_prog, 10178 struct bpf_prog *old_prog) 10179 { 10180 struct bpf_xdp_link *xdp_link = container_of(link, struct bpf_xdp_link, link); 10181 enum bpf_xdp_mode mode; 10182 bpf_op_t bpf_op; 10183 int err = 0; 10184 10185 rtnl_lock(); 10186 10187 /* link might have been auto-released already, so fail */ 10188 if (!xdp_link->dev) { 10189 err = -ENOLINK; 10190 goto out_unlock; 10191 } 10192 10193 if (old_prog && link->prog != old_prog) { 10194 err = -EPERM; 10195 goto out_unlock; 10196 } 10197 old_prog = link->prog; 10198 if (old_prog->type != new_prog->type || 10199 old_prog->expected_attach_type != new_prog->expected_attach_type) { 10200 err = -EINVAL; 10201 goto out_unlock; 10202 } 10203 10204 if (old_prog == new_prog) { 10205 /* no-op, don't disturb drivers */ 10206 bpf_prog_put(new_prog); 10207 goto out_unlock; 10208 } 10209 10210 netdev_lock_ops(xdp_link->dev); 10211 mode = dev_xdp_mode(xdp_link->dev, xdp_link->flags); 10212 bpf_op = dev_xdp_bpf_op(xdp_link->dev, mode); 10213 err = dev_xdp_install(xdp_link->dev, mode, bpf_op, NULL, 10214 xdp_link->flags, new_prog); 10215 netdev_unlock_ops(xdp_link->dev); 10216 if (err) 10217 goto out_unlock; 10218 10219 old_prog = xchg(&link->prog, new_prog); 10220 bpf_prog_put(old_prog); 10221 10222 out_unlock: 10223 rtnl_unlock(); 10224 return err; 10225 } 10226 10227 static const struct bpf_link_ops bpf_xdp_link_lops = { 10228 .release = bpf_xdp_link_release, 10229 .dealloc = bpf_xdp_link_dealloc, 10230 .detach = bpf_xdp_link_detach, 10231 .show_fdinfo = bpf_xdp_link_show_fdinfo, 10232 .fill_link_info = bpf_xdp_link_fill_link_info, 10233 .update_prog = bpf_xdp_link_update, 10234 }; 10235 10236 int bpf_xdp_link_attach(const union bpf_attr *attr, struct bpf_prog *prog) 10237 { 10238 struct net *net = current->nsproxy->net_ns; 10239 struct bpf_link_primer link_primer; 10240 struct netlink_ext_ack extack = {}; 10241 struct bpf_xdp_link *link; 10242 struct net_device *dev; 10243 int err, fd; 10244 10245 rtnl_lock(); 10246 dev = dev_get_by_index(net, attr->link_create.target_ifindex); 10247 if (!dev) { 10248 rtnl_unlock(); 10249 return -EINVAL; 10250 } 10251 10252 link = kzalloc(sizeof(*link), GFP_USER); 10253 if (!link) { 10254 err = -ENOMEM; 10255 goto unlock; 10256 } 10257 10258 bpf_link_init(&link->link, BPF_LINK_TYPE_XDP, &bpf_xdp_link_lops, prog); 10259 link->dev = dev; 10260 link->flags = attr->link_create.flags; 10261 10262 err = bpf_link_prime(&link->link, &link_primer); 10263 if (err) { 10264 kfree(link); 10265 goto unlock; 10266 } 10267 10268 err = dev_xdp_attach_link(dev, &extack, link); 10269 rtnl_unlock(); 10270 10271 if (err) { 10272 link->dev = NULL; 10273 bpf_link_cleanup(&link_primer); 10274 trace_bpf_xdp_link_attach_failed(extack._msg); 10275 goto out_put_dev; 10276 } 10277 10278 fd = bpf_link_settle(&link_primer); 10279 /* link itself doesn't hold dev's refcnt to not complicate shutdown */ 10280 dev_put(dev); 10281 return fd; 10282 10283 unlock: 10284 rtnl_unlock(); 10285 10286 out_put_dev: 10287 dev_put(dev); 10288 return err; 10289 } 10290 10291 /** 10292 * dev_change_xdp_fd - set or clear a bpf program for a device rx path 10293 * @dev: device 10294 * @extack: netlink extended ack 10295 * @fd: new program fd or negative value to clear 10296 * @expected_fd: old program fd that userspace expects to replace or clear 10297 * @flags: xdp-related flags 10298 * 10299 * Set or clear a bpf program for a device 10300 */ 10301 int dev_change_xdp_fd(struct net_device *dev, struct netlink_ext_ack *extack, 10302 int fd, int expected_fd, u32 flags) 10303 { 10304 enum bpf_xdp_mode mode = dev_xdp_mode(dev, flags); 10305 struct bpf_prog *new_prog = NULL, *old_prog = NULL; 10306 int err; 10307 10308 ASSERT_RTNL(); 10309 10310 if (fd >= 0) { 10311 new_prog = bpf_prog_get_type_dev(fd, BPF_PROG_TYPE_XDP, 10312 mode != XDP_MODE_SKB); 10313 if (IS_ERR(new_prog)) 10314 return PTR_ERR(new_prog); 10315 } 10316 10317 if (expected_fd >= 0) { 10318 old_prog = bpf_prog_get_type_dev(expected_fd, BPF_PROG_TYPE_XDP, 10319 mode != XDP_MODE_SKB); 10320 if (IS_ERR(old_prog)) { 10321 err = PTR_ERR(old_prog); 10322 old_prog = NULL; 10323 goto err_out; 10324 } 10325 } 10326 10327 err = dev_xdp_attach(dev, extack, NULL, new_prog, old_prog, flags); 10328 10329 err_out: 10330 if (err && new_prog) 10331 bpf_prog_put(new_prog); 10332 if (old_prog) 10333 bpf_prog_put(old_prog); 10334 return err; 10335 } 10336 10337 u32 dev_get_min_mp_channel_count(const struct net_device *dev) 10338 { 10339 int i; 10340 10341 ASSERT_RTNL(); 10342 10343 for (i = dev->real_num_rx_queues - 1; i >= 0; i--) 10344 if (dev->_rx[i].mp_params.mp_priv) 10345 /* The channel count is the idx plus 1. */ 10346 return i + 1; 10347 10348 return 0; 10349 } 10350 10351 /** 10352 * dev_index_reserve() - allocate an ifindex in a namespace 10353 * @net: the applicable net namespace 10354 * @ifindex: requested ifindex, pass %0 to get one allocated 10355 * 10356 * Allocate a ifindex for a new device. Caller must either use the ifindex 10357 * to store the device (via list_netdevice()) or call dev_index_release() 10358 * to give the index up. 10359 * 10360 * Return: a suitable unique value for a new device interface number or -errno. 10361 */ 10362 static int dev_index_reserve(struct net *net, u32 ifindex) 10363 { 10364 int err; 10365 10366 if (ifindex > INT_MAX) { 10367 DEBUG_NET_WARN_ON_ONCE(1); 10368 return -EINVAL; 10369 } 10370 10371 if (!ifindex) 10372 err = xa_alloc_cyclic(&net->dev_by_index, &ifindex, NULL, 10373 xa_limit_31b, &net->ifindex, GFP_KERNEL); 10374 else 10375 err = xa_insert(&net->dev_by_index, ifindex, NULL, GFP_KERNEL); 10376 if (err < 0) 10377 return err; 10378 10379 return ifindex; 10380 } 10381 10382 static void dev_index_release(struct net *net, int ifindex) 10383 { 10384 /* Expect only unused indexes, unlist_netdevice() removes the used */ 10385 WARN_ON(xa_erase(&net->dev_by_index, ifindex)); 10386 } 10387 10388 static bool from_cleanup_net(void) 10389 { 10390 #ifdef CONFIG_NET_NS 10391 return current == cleanup_net_task; 10392 #else 10393 return false; 10394 #endif 10395 } 10396 10397 /* Delayed registration/unregisteration */ 10398 LIST_HEAD(net_todo_list); 10399 DECLARE_WAIT_QUEUE_HEAD(netdev_unregistering_wq); 10400 atomic_t dev_unreg_count = ATOMIC_INIT(0); 10401 10402 static void net_set_todo(struct net_device *dev) 10403 { 10404 list_add_tail(&dev->todo_list, &net_todo_list); 10405 } 10406 10407 static netdev_features_t netdev_sync_upper_features(struct net_device *lower, 10408 struct net_device *upper, netdev_features_t features) 10409 { 10410 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES; 10411 netdev_features_t feature; 10412 int feature_bit; 10413 10414 for_each_netdev_feature(upper_disables, feature_bit) { 10415 feature = __NETIF_F_BIT(feature_bit); 10416 if (!(upper->wanted_features & feature) 10417 && (features & feature)) { 10418 netdev_dbg(lower, "Dropping feature %pNF, upper dev %s has it off.\n", 10419 &feature, upper->name); 10420 features &= ~feature; 10421 } 10422 } 10423 10424 return features; 10425 } 10426 10427 static void netdev_sync_lower_features(struct net_device *upper, 10428 struct net_device *lower, netdev_features_t features) 10429 { 10430 netdev_features_t upper_disables = NETIF_F_UPPER_DISABLES; 10431 netdev_features_t feature; 10432 int feature_bit; 10433 10434 for_each_netdev_feature(upper_disables, feature_bit) { 10435 feature = __NETIF_F_BIT(feature_bit); 10436 if (!(features & feature) && (lower->features & feature)) { 10437 netdev_dbg(upper, "Disabling feature %pNF on lower dev %s.\n", 10438 &feature, lower->name); 10439 lower->wanted_features &= ~feature; 10440 __netdev_update_features(lower); 10441 10442 if (unlikely(lower->features & feature)) 10443 netdev_WARN(upper, "failed to disable %pNF on %s!\n", 10444 &feature, lower->name); 10445 else 10446 netdev_features_change(lower); 10447 } 10448 } 10449 } 10450 10451 static bool netdev_has_ip_or_hw_csum(netdev_features_t features) 10452 { 10453 netdev_features_t ip_csum_mask = NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM; 10454 bool ip_csum = (features & ip_csum_mask) == ip_csum_mask; 10455 bool hw_csum = features & NETIF_F_HW_CSUM; 10456 10457 return ip_csum || hw_csum; 10458 } 10459 10460 static netdev_features_t netdev_fix_features(struct net_device *dev, 10461 netdev_features_t features) 10462 { 10463 /* Fix illegal checksum combinations */ 10464 if ((features & NETIF_F_HW_CSUM) && 10465 (features & (NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM))) { 10466 netdev_warn(dev, "mixed HW and IP checksum settings.\n"); 10467 features &= ~(NETIF_F_IP_CSUM|NETIF_F_IPV6_CSUM); 10468 } 10469 10470 /* TSO requires that SG is present as well. */ 10471 if ((features & NETIF_F_ALL_TSO) && !(features & NETIF_F_SG)) { 10472 netdev_dbg(dev, "Dropping TSO features since no SG feature.\n"); 10473 features &= ~NETIF_F_ALL_TSO; 10474 } 10475 10476 if ((features & NETIF_F_TSO) && !(features & NETIF_F_HW_CSUM) && 10477 !(features & NETIF_F_IP_CSUM)) { 10478 netdev_dbg(dev, "Dropping TSO features since no CSUM feature.\n"); 10479 features &= ~NETIF_F_TSO; 10480 features &= ~NETIF_F_TSO_ECN; 10481 } 10482 10483 if ((features & NETIF_F_TSO6) && !(features & NETIF_F_HW_CSUM) && 10484 !(features & NETIF_F_IPV6_CSUM)) { 10485 netdev_dbg(dev, "Dropping TSO6 features since no CSUM feature.\n"); 10486 features &= ~NETIF_F_TSO6; 10487 } 10488 10489 /* TSO with IPv4 ID mangling requires IPv4 TSO be enabled */ 10490 if ((features & NETIF_F_TSO_MANGLEID) && !(features & NETIF_F_TSO)) 10491 features &= ~NETIF_F_TSO_MANGLEID; 10492 10493 /* TSO ECN requires that TSO is present as well. */ 10494 if ((features & NETIF_F_ALL_TSO) == NETIF_F_TSO_ECN) 10495 features &= ~NETIF_F_TSO_ECN; 10496 10497 /* Software GSO depends on SG. */ 10498 if ((features & NETIF_F_GSO) && !(features & NETIF_F_SG)) { 10499 netdev_dbg(dev, "Dropping NETIF_F_GSO since no SG feature.\n"); 10500 features &= ~NETIF_F_GSO; 10501 } 10502 10503 /* GSO partial features require GSO partial be set */ 10504 if ((features & dev->gso_partial_features) && 10505 !(features & NETIF_F_GSO_PARTIAL)) { 10506 netdev_dbg(dev, 10507 "Dropping partially supported GSO features since no GSO partial.\n"); 10508 features &= ~dev->gso_partial_features; 10509 } 10510 10511 if (!(features & NETIF_F_RXCSUM)) { 10512 /* NETIF_F_GRO_HW implies doing RXCSUM since every packet 10513 * successfully merged by hardware must also have the 10514 * checksum verified by hardware. If the user does not 10515 * want to enable RXCSUM, logically, we should disable GRO_HW. 10516 */ 10517 if (features & NETIF_F_GRO_HW) { 10518 netdev_dbg(dev, "Dropping NETIF_F_GRO_HW since no RXCSUM feature.\n"); 10519 features &= ~NETIF_F_GRO_HW; 10520 } 10521 } 10522 10523 /* LRO/HW-GRO features cannot be combined with RX-FCS */ 10524 if (features & NETIF_F_RXFCS) { 10525 if (features & NETIF_F_LRO) { 10526 netdev_dbg(dev, "Dropping LRO feature since RX-FCS is requested.\n"); 10527 features &= ~NETIF_F_LRO; 10528 } 10529 10530 if (features & NETIF_F_GRO_HW) { 10531 netdev_dbg(dev, "Dropping HW-GRO feature since RX-FCS is requested.\n"); 10532 features &= ~NETIF_F_GRO_HW; 10533 } 10534 } 10535 10536 if ((features & NETIF_F_GRO_HW) && (features & NETIF_F_LRO)) { 10537 netdev_dbg(dev, "Dropping LRO feature since HW-GRO is requested.\n"); 10538 features &= ~NETIF_F_LRO; 10539 } 10540 10541 if ((features & NETIF_F_HW_TLS_TX) && !netdev_has_ip_or_hw_csum(features)) { 10542 netdev_dbg(dev, "Dropping TLS TX HW offload feature since no CSUM feature.\n"); 10543 features &= ~NETIF_F_HW_TLS_TX; 10544 } 10545 10546 if ((features & NETIF_F_HW_TLS_RX) && !(features & NETIF_F_RXCSUM)) { 10547 netdev_dbg(dev, "Dropping TLS RX HW offload feature since no RXCSUM feature.\n"); 10548 features &= ~NETIF_F_HW_TLS_RX; 10549 } 10550 10551 if ((features & NETIF_F_GSO_UDP_L4) && !netdev_has_ip_or_hw_csum(features)) { 10552 netdev_dbg(dev, "Dropping USO feature since no CSUM feature.\n"); 10553 features &= ~NETIF_F_GSO_UDP_L4; 10554 } 10555 10556 return features; 10557 } 10558 10559 int __netdev_update_features(struct net_device *dev) 10560 { 10561 struct net_device *upper, *lower; 10562 netdev_features_t features; 10563 struct list_head *iter; 10564 int err = -1; 10565 10566 ASSERT_RTNL(); 10567 netdev_ops_assert_locked(dev); 10568 10569 features = netdev_get_wanted_features(dev); 10570 10571 if (dev->netdev_ops->ndo_fix_features) 10572 features = dev->netdev_ops->ndo_fix_features(dev, features); 10573 10574 /* driver might be less strict about feature dependencies */ 10575 features = netdev_fix_features(dev, features); 10576 10577 /* some features can't be enabled if they're off on an upper device */ 10578 netdev_for_each_upper_dev_rcu(dev, upper, iter) 10579 features = netdev_sync_upper_features(dev, upper, features); 10580 10581 if (dev->features == features) 10582 goto sync_lower; 10583 10584 netdev_dbg(dev, "Features changed: %pNF -> %pNF\n", 10585 &dev->features, &features); 10586 10587 if (dev->netdev_ops->ndo_set_features) 10588 err = dev->netdev_ops->ndo_set_features(dev, features); 10589 else 10590 err = 0; 10591 10592 if (unlikely(err < 0)) { 10593 netdev_err(dev, 10594 "set_features() failed (%d); wanted %pNF, left %pNF\n", 10595 err, &features, &dev->features); 10596 /* return non-0 since some features might have changed and 10597 * it's better to fire a spurious notification than miss it 10598 */ 10599 return -1; 10600 } 10601 10602 sync_lower: 10603 /* some features must be disabled on lower devices when disabled 10604 * on an upper device (think: bonding master or bridge) 10605 */ 10606 netdev_for_each_lower_dev(dev, lower, iter) 10607 netdev_sync_lower_features(dev, lower, features); 10608 10609 if (!err) { 10610 netdev_features_t diff = features ^ dev->features; 10611 10612 if (diff & NETIF_F_RX_UDP_TUNNEL_PORT) { 10613 /* udp_tunnel_{get,drop}_rx_info both need 10614 * NETIF_F_RX_UDP_TUNNEL_PORT enabled on the 10615 * device, or they won't do anything. 10616 * Thus we need to update dev->features 10617 * *before* calling udp_tunnel_get_rx_info, 10618 * but *after* calling udp_tunnel_drop_rx_info. 10619 */ 10620 if (features & NETIF_F_RX_UDP_TUNNEL_PORT) { 10621 dev->features = features; 10622 udp_tunnel_get_rx_info(dev); 10623 } else { 10624 udp_tunnel_drop_rx_info(dev); 10625 } 10626 } 10627 10628 if (diff & NETIF_F_HW_VLAN_CTAG_FILTER) { 10629 if (features & NETIF_F_HW_VLAN_CTAG_FILTER) { 10630 dev->features = features; 10631 err |= vlan_get_rx_ctag_filter_info(dev); 10632 } else { 10633 vlan_drop_rx_ctag_filter_info(dev); 10634 } 10635 } 10636 10637 if (diff & NETIF_F_HW_VLAN_STAG_FILTER) { 10638 if (features & NETIF_F_HW_VLAN_STAG_FILTER) { 10639 dev->features = features; 10640 err |= vlan_get_rx_stag_filter_info(dev); 10641 } else { 10642 vlan_drop_rx_stag_filter_info(dev); 10643 } 10644 } 10645 10646 dev->features = features; 10647 } 10648 10649 return err < 0 ? 0 : 1; 10650 } 10651 10652 /** 10653 * netdev_update_features - recalculate device features 10654 * @dev: the device to check 10655 * 10656 * Recalculate dev->features set and send notifications if it 10657 * has changed. Should be called after driver or hardware dependent 10658 * conditions might have changed that influence the features. 10659 */ 10660 void netdev_update_features(struct net_device *dev) 10661 { 10662 if (__netdev_update_features(dev)) 10663 netdev_features_change(dev); 10664 } 10665 EXPORT_SYMBOL(netdev_update_features); 10666 10667 /** 10668 * netdev_change_features - recalculate device features 10669 * @dev: the device to check 10670 * 10671 * Recalculate dev->features set and send notifications even 10672 * if they have not changed. Should be called instead of 10673 * netdev_update_features() if also dev->vlan_features might 10674 * have changed to allow the changes to be propagated to stacked 10675 * VLAN devices. 10676 */ 10677 void netdev_change_features(struct net_device *dev) 10678 { 10679 __netdev_update_features(dev); 10680 netdev_features_change(dev); 10681 } 10682 EXPORT_SYMBOL(netdev_change_features); 10683 10684 /** 10685 * netif_stacked_transfer_operstate - transfer operstate 10686 * @rootdev: the root or lower level device to transfer state from 10687 * @dev: the device to transfer operstate to 10688 * 10689 * Transfer operational state from root to device. This is normally 10690 * called when a stacking relationship exists between the root 10691 * device and the device(a leaf device). 10692 */ 10693 void netif_stacked_transfer_operstate(const struct net_device *rootdev, 10694 struct net_device *dev) 10695 { 10696 if (rootdev->operstate == IF_OPER_DORMANT) 10697 netif_dormant_on(dev); 10698 else 10699 netif_dormant_off(dev); 10700 10701 if (rootdev->operstate == IF_OPER_TESTING) 10702 netif_testing_on(dev); 10703 else 10704 netif_testing_off(dev); 10705 10706 if (netif_carrier_ok(rootdev)) 10707 netif_carrier_on(dev); 10708 else 10709 netif_carrier_off(dev); 10710 } 10711 EXPORT_SYMBOL(netif_stacked_transfer_operstate); 10712 10713 static int netif_alloc_rx_queues(struct net_device *dev) 10714 { 10715 unsigned int i, count = dev->num_rx_queues; 10716 struct netdev_rx_queue *rx; 10717 size_t sz = count * sizeof(*rx); 10718 int err = 0; 10719 10720 BUG_ON(count < 1); 10721 10722 rx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 10723 if (!rx) 10724 return -ENOMEM; 10725 10726 dev->_rx = rx; 10727 10728 for (i = 0; i < count; i++) { 10729 rx[i].dev = dev; 10730 10731 /* XDP RX-queue setup */ 10732 err = xdp_rxq_info_reg(&rx[i].xdp_rxq, dev, i, 0); 10733 if (err < 0) 10734 goto err_rxq_info; 10735 } 10736 return 0; 10737 10738 err_rxq_info: 10739 /* Rollback successful reg's and free other resources */ 10740 while (i--) 10741 xdp_rxq_info_unreg(&rx[i].xdp_rxq); 10742 kvfree(dev->_rx); 10743 dev->_rx = NULL; 10744 return err; 10745 } 10746 10747 static void netif_free_rx_queues(struct net_device *dev) 10748 { 10749 unsigned int i, count = dev->num_rx_queues; 10750 10751 /* netif_alloc_rx_queues alloc failed, resources have been unreg'ed */ 10752 if (!dev->_rx) 10753 return; 10754 10755 for (i = 0; i < count; i++) 10756 xdp_rxq_info_unreg(&dev->_rx[i].xdp_rxq); 10757 10758 kvfree(dev->_rx); 10759 } 10760 10761 static void netdev_init_one_queue(struct net_device *dev, 10762 struct netdev_queue *queue, void *_unused) 10763 { 10764 /* Initialize queue lock */ 10765 spin_lock_init(&queue->_xmit_lock); 10766 netdev_set_xmit_lockdep_class(&queue->_xmit_lock, dev->type); 10767 queue->xmit_lock_owner = -1; 10768 netdev_queue_numa_node_write(queue, NUMA_NO_NODE); 10769 queue->dev = dev; 10770 #ifdef CONFIG_BQL 10771 dql_init(&queue->dql, HZ); 10772 #endif 10773 } 10774 10775 static void netif_free_tx_queues(struct net_device *dev) 10776 { 10777 kvfree(dev->_tx); 10778 } 10779 10780 static int netif_alloc_netdev_queues(struct net_device *dev) 10781 { 10782 unsigned int count = dev->num_tx_queues; 10783 struct netdev_queue *tx; 10784 size_t sz = count * sizeof(*tx); 10785 10786 if (count < 1 || count > 0xffff) 10787 return -EINVAL; 10788 10789 tx = kvzalloc(sz, GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 10790 if (!tx) 10791 return -ENOMEM; 10792 10793 dev->_tx = tx; 10794 10795 netdev_for_each_tx_queue(dev, netdev_init_one_queue, NULL); 10796 spin_lock_init(&dev->tx_global_lock); 10797 10798 return 0; 10799 } 10800 10801 void netif_tx_stop_all_queues(struct net_device *dev) 10802 { 10803 unsigned int i; 10804 10805 for (i = 0; i < dev->num_tx_queues; i++) { 10806 struct netdev_queue *txq = netdev_get_tx_queue(dev, i); 10807 10808 netif_tx_stop_queue(txq); 10809 } 10810 } 10811 EXPORT_SYMBOL(netif_tx_stop_all_queues); 10812 10813 static int netdev_do_alloc_pcpu_stats(struct net_device *dev) 10814 { 10815 void __percpu *v; 10816 10817 /* Drivers implementing ndo_get_peer_dev must support tstat 10818 * accounting, so that skb_do_redirect() can bump the dev's 10819 * RX stats upon network namespace switch. 10820 */ 10821 if (dev->netdev_ops->ndo_get_peer_dev && 10822 dev->pcpu_stat_type != NETDEV_PCPU_STAT_TSTATS) 10823 return -EOPNOTSUPP; 10824 10825 switch (dev->pcpu_stat_type) { 10826 case NETDEV_PCPU_STAT_NONE: 10827 return 0; 10828 case NETDEV_PCPU_STAT_LSTATS: 10829 v = dev->lstats = netdev_alloc_pcpu_stats(struct pcpu_lstats); 10830 break; 10831 case NETDEV_PCPU_STAT_TSTATS: 10832 v = dev->tstats = netdev_alloc_pcpu_stats(struct pcpu_sw_netstats); 10833 break; 10834 case NETDEV_PCPU_STAT_DSTATS: 10835 v = dev->dstats = netdev_alloc_pcpu_stats(struct pcpu_dstats); 10836 break; 10837 default: 10838 return -EINVAL; 10839 } 10840 10841 return v ? 0 : -ENOMEM; 10842 } 10843 10844 static void netdev_do_free_pcpu_stats(struct net_device *dev) 10845 { 10846 switch (dev->pcpu_stat_type) { 10847 case NETDEV_PCPU_STAT_NONE: 10848 return; 10849 case NETDEV_PCPU_STAT_LSTATS: 10850 free_percpu(dev->lstats); 10851 break; 10852 case NETDEV_PCPU_STAT_TSTATS: 10853 free_percpu(dev->tstats); 10854 break; 10855 case NETDEV_PCPU_STAT_DSTATS: 10856 free_percpu(dev->dstats); 10857 break; 10858 } 10859 } 10860 10861 static void netdev_free_phy_link_topology(struct net_device *dev) 10862 { 10863 struct phy_link_topology *topo = dev->link_topo; 10864 10865 if (IS_ENABLED(CONFIG_PHYLIB) && topo) { 10866 xa_destroy(&topo->phys); 10867 kfree(topo); 10868 dev->link_topo = NULL; 10869 } 10870 } 10871 10872 /** 10873 * register_netdevice() - register a network device 10874 * @dev: device to register 10875 * 10876 * Take a prepared network device structure and make it externally accessible. 10877 * A %NETDEV_REGISTER message is sent to the netdev notifier chain. 10878 * Callers must hold the rtnl lock - you may want register_netdev() 10879 * instead of this. 10880 */ 10881 int register_netdevice(struct net_device *dev) 10882 { 10883 int ret; 10884 struct net *net = dev_net(dev); 10885 10886 BUILD_BUG_ON(sizeof(netdev_features_t) * BITS_PER_BYTE < 10887 NETDEV_FEATURE_COUNT); 10888 BUG_ON(dev_boot_phase); 10889 ASSERT_RTNL(); 10890 10891 might_sleep(); 10892 10893 /* When net_device's are persistent, this will be fatal. */ 10894 BUG_ON(dev->reg_state != NETREG_UNINITIALIZED); 10895 BUG_ON(!net); 10896 10897 ret = ethtool_check_ops(dev->ethtool_ops); 10898 if (ret) 10899 return ret; 10900 10901 /* rss ctx ID 0 is reserved for the default context, start from 1 */ 10902 xa_init_flags(&dev->ethtool->rss_ctx, XA_FLAGS_ALLOC1); 10903 mutex_init(&dev->ethtool->rss_lock); 10904 10905 spin_lock_init(&dev->addr_list_lock); 10906 netdev_set_addr_lockdep_class(dev); 10907 10908 ret = dev_get_valid_name(net, dev, dev->name); 10909 if (ret < 0) 10910 goto out; 10911 10912 ret = -ENOMEM; 10913 dev->name_node = netdev_name_node_head_alloc(dev); 10914 if (!dev->name_node) 10915 goto out; 10916 10917 /* Init, if this function is available */ 10918 if (dev->netdev_ops->ndo_init) { 10919 ret = dev->netdev_ops->ndo_init(dev); 10920 if (ret) { 10921 if (ret > 0) 10922 ret = -EIO; 10923 goto err_free_name; 10924 } 10925 } 10926 10927 if (((dev->hw_features | dev->features) & 10928 NETIF_F_HW_VLAN_CTAG_FILTER) && 10929 (!dev->netdev_ops->ndo_vlan_rx_add_vid || 10930 !dev->netdev_ops->ndo_vlan_rx_kill_vid)) { 10931 netdev_WARN(dev, "Buggy VLAN acceleration in driver!\n"); 10932 ret = -EINVAL; 10933 goto err_uninit; 10934 } 10935 10936 ret = netdev_do_alloc_pcpu_stats(dev); 10937 if (ret) 10938 goto err_uninit; 10939 10940 ret = dev_index_reserve(net, dev->ifindex); 10941 if (ret < 0) 10942 goto err_free_pcpu; 10943 dev->ifindex = ret; 10944 10945 /* Transfer changeable features to wanted_features and enable 10946 * software offloads (GSO and GRO). 10947 */ 10948 dev->hw_features |= (NETIF_F_SOFT_FEATURES | NETIF_F_SOFT_FEATURES_OFF); 10949 dev->features |= NETIF_F_SOFT_FEATURES; 10950 10951 if (dev->udp_tunnel_nic_info) { 10952 dev->features |= NETIF_F_RX_UDP_TUNNEL_PORT; 10953 dev->hw_features |= NETIF_F_RX_UDP_TUNNEL_PORT; 10954 } 10955 10956 dev->wanted_features = dev->features & dev->hw_features; 10957 10958 if (!(dev->flags & IFF_LOOPBACK)) 10959 dev->hw_features |= NETIF_F_NOCACHE_COPY; 10960 10961 /* If IPv4 TCP segmentation offload is supported we should also 10962 * allow the device to enable segmenting the frame with the option 10963 * of ignoring a static IP ID value. This doesn't enable the 10964 * feature itself but allows the user to enable it later. 10965 */ 10966 if (dev->hw_features & NETIF_F_TSO) 10967 dev->hw_features |= NETIF_F_TSO_MANGLEID; 10968 if (dev->vlan_features & NETIF_F_TSO) 10969 dev->vlan_features |= NETIF_F_TSO_MANGLEID; 10970 if (dev->mpls_features & NETIF_F_TSO) 10971 dev->mpls_features |= NETIF_F_TSO_MANGLEID; 10972 if (dev->hw_enc_features & NETIF_F_TSO) 10973 dev->hw_enc_features |= NETIF_F_TSO_MANGLEID; 10974 10975 /* Make NETIF_F_HIGHDMA inheritable to VLAN devices. 10976 */ 10977 dev->vlan_features |= NETIF_F_HIGHDMA; 10978 10979 /* Make NETIF_F_SG inheritable to tunnel devices. 10980 */ 10981 dev->hw_enc_features |= NETIF_F_SG | NETIF_F_GSO_PARTIAL; 10982 10983 /* Make NETIF_F_SG inheritable to MPLS. 10984 */ 10985 dev->mpls_features |= NETIF_F_SG; 10986 10987 ret = call_netdevice_notifiers(NETDEV_POST_INIT, dev); 10988 ret = notifier_to_errno(ret); 10989 if (ret) 10990 goto err_ifindex_release; 10991 10992 ret = netdev_register_kobject(dev); 10993 10994 netdev_lock(dev); 10995 WRITE_ONCE(dev->reg_state, ret ? NETREG_UNREGISTERED : NETREG_REGISTERED); 10996 netdev_unlock(dev); 10997 10998 if (ret) 10999 goto err_uninit_notify; 11000 11001 netdev_lock_ops(dev); 11002 __netdev_update_features(dev); 11003 netdev_unlock_ops(dev); 11004 11005 /* 11006 * Default initial state at registry is that the 11007 * device is present. 11008 */ 11009 11010 set_bit(__LINK_STATE_PRESENT, &dev->state); 11011 11012 linkwatch_init_dev(dev); 11013 11014 dev_init_scheduler(dev); 11015 11016 netdev_hold(dev, &dev->dev_registered_tracker, GFP_KERNEL); 11017 list_netdevice(dev); 11018 11019 add_device_randomness(dev->dev_addr, dev->addr_len); 11020 11021 /* If the device has permanent device address, driver should 11022 * set dev_addr and also addr_assign_type should be set to 11023 * NET_ADDR_PERM (default value). 11024 */ 11025 if (dev->addr_assign_type == NET_ADDR_PERM) 11026 memcpy(dev->perm_addr, dev->dev_addr, dev->addr_len); 11027 11028 /* Notify protocols, that a new device appeared. */ 11029 ret = call_netdevice_notifiers(NETDEV_REGISTER, dev); 11030 ret = notifier_to_errno(ret); 11031 if (ret) { 11032 /* Expect explicit free_netdev() on failure */ 11033 dev->needs_free_netdev = false; 11034 unregister_netdevice_queue(dev, NULL); 11035 goto out; 11036 } 11037 /* 11038 * Prevent userspace races by waiting until the network 11039 * device is fully setup before sending notifications. 11040 */ 11041 if (!dev->rtnl_link_ops || 11042 dev->rtnl_link_state == RTNL_LINK_INITIALIZED) 11043 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL, 0, NULL); 11044 11045 out: 11046 return ret; 11047 11048 err_uninit_notify: 11049 call_netdevice_notifiers(NETDEV_PRE_UNINIT, dev); 11050 err_ifindex_release: 11051 dev_index_release(net, dev->ifindex); 11052 err_free_pcpu: 11053 netdev_do_free_pcpu_stats(dev); 11054 err_uninit: 11055 if (dev->netdev_ops->ndo_uninit) 11056 dev->netdev_ops->ndo_uninit(dev); 11057 if (dev->priv_destructor) 11058 dev->priv_destructor(dev); 11059 err_free_name: 11060 netdev_name_node_free(dev->name_node); 11061 goto out; 11062 } 11063 EXPORT_SYMBOL(register_netdevice); 11064 11065 /* Initialize the core of a dummy net device. 11066 * The setup steps dummy netdevs need which normal netdevs get by going 11067 * through register_netdevice(). 11068 */ 11069 static void init_dummy_netdev(struct net_device *dev) 11070 { 11071 /* make sure we BUG if trying to hit standard 11072 * register/unregister code path 11073 */ 11074 dev->reg_state = NETREG_DUMMY; 11075 11076 /* a dummy interface is started by default */ 11077 set_bit(__LINK_STATE_PRESENT, &dev->state); 11078 set_bit(__LINK_STATE_START, &dev->state); 11079 11080 /* Note : We dont allocate pcpu_refcnt for dummy devices, 11081 * because users of this 'device' dont need to change 11082 * its refcount. 11083 */ 11084 } 11085 11086 /** 11087 * register_netdev - register a network device 11088 * @dev: device to register 11089 * 11090 * Take a completed network device structure and add it to the kernel 11091 * interfaces. A %NETDEV_REGISTER message is sent to the netdev notifier 11092 * chain. 0 is returned on success. A negative errno code is returned 11093 * on a failure to set up the device, or if the name is a duplicate. 11094 * 11095 * This is a wrapper around register_netdevice that takes the rtnl semaphore 11096 * and expands the device name if you passed a format string to 11097 * alloc_netdev. 11098 */ 11099 int register_netdev(struct net_device *dev) 11100 { 11101 struct net *net = dev_net(dev); 11102 int err; 11103 11104 if (rtnl_net_lock_killable(net)) 11105 return -EINTR; 11106 11107 err = register_netdevice(dev); 11108 11109 rtnl_net_unlock(net); 11110 11111 return err; 11112 } 11113 EXPORT_SYMBOL(register_netdev); 11114 11115 int netdev_refcnt_read(const struct net_device *dev) 11116 { 11117 #ifdef CONFIG_PCPU_DEV_REFCNT 11118 int i, refcnt = 0; 11119 11120 for_each_possible_cpu(i) 11121 refcnt += *per_cpu_ptr(dev->pcpu_refcnt, i); 11122 return refcnt; 11123 #else 11124 return refcount_read(&dev->dev_refcnt); 11125 #endif 11126 } 11127 EXPORT_SYMBOL(netdev_refcnt_read); 11128 11129 int netdev_unregister_timeout_secs __read_mostly = 10; 11130 11131 #define WAIT_REFS_MIN_MSECS 1 11132 #define WAIT_REFS_MAX_MSECS 250 11133 /** 11134 * netdev_wait_allrefs_any - wait until all references are gone. 11135 * @list: list of net_devices to wait on 11136 * 11137 * This is called when unregistering network devices. 11138 * 11139 * Any protocol or device that holds a reference should register 11140 * for netdevice notification, and cleanup and put back the 11141 * reference if they receive an UNREGISTER event. 11142 * We can get stuck here if buggy protocols don't correctly 11143 * call dev_put. 11144 */ 11145 static struct net_device *netdev_wait_allrefs_any(struct list_head *list) 11146 { 11147 unsigned long rebroadcast_time, warning_time; 11148 struct net_device *dev; 11149 int wait = 0; 11150 11151 rebroadcast_time = warning_time = jiffies; 11152 11153 list_for_each_entry(dev, list, todo_list) 11154 if (netdev_refcnt_read(dev) == 1) 11155 return dev; 11156 11157 while (true) { 11158 if (time_after(jiffies, rebroadcast_time + 1 * HZ)) { 11159 rtnl_lock(); 11160 11161 /* Rebroadcast unregister notification */ 11162 list_for_each_entry(dev, list, todo_list) 11163 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 11164 11165 __rtnl_unlock(); 11166 rcu_barrier(); 11167 rtnl_lock(); 11168 11169 list_for_each_entry(dev, list, todo_list) 11170 if (test_bit(__LINK_STATE_LINKWATCH_PENDING, 11171 &dev->state)) { 11172 /* We must not have linkwatch events 11173 * pending on unregister. If this 11174 * happens, we simply run the queue 11175 * unscheduled, resulting in a noop 11176 * for this device. 11177 */ 11178 linkwatch_run_queue(); 11179 break; 11180 } 11181 11182 __rtnl_unlock(); 11183 11184 rebroadcast_time = jiffies; 11185 } 11186 11187 rcu_barrier(); 11188 11189 if (!wait) { 11190 wait = WAIT_REFS_MIN_MSECS; 11191 } else { 11192 msleep(wait); 11193 wait = min(wait << 1, WAIT_REFS_MAX_MSECS); 11194 } 11195 11196 list_for_each_entry(dev, list, todo_list) 11197 if (netdev_refcnt_read(dev) == 1) 11198 return dev; 11199 11200 if (time_after(jiffies, warning_time + 11201 READ_ONCE(netdev_unregister_timeout_secs) * HZ)) { 11202 list_for_each_entry(dev, list, todo_list) { 11203 pr_emerg("unregister_netdevice: waiting for %s to become free. Usage count = %d\n", 11204 dev->name, netdev_refcnt_read(dev)); 11205 ref_tracker_dir_print(&dev->refcnt_tracker, 10); 11206 } 11207 11208 warning_time = jiffies; 11209 } 11210 } 11211 } 11212 11213 /* The sequence is: 11214 * 11215 * rtnl_lock(); 11216 * ... 11217 * register_netdevice(x1); 11218 * register_netdevice(x2); 11219 * ... 11220 * unregister_netdevice(y1); 11221 * unregister_netdevice(y2); 11222 * ... 11223 * rtnl_unlock(); 11224 * free_netdev(y1); 11225 * free_netdev(y2); 11226 * 11227 * We are invoked by rtnl_unlock(). 11228 * This allows us to deal with problems: 11229 * 1) We can delete sysfs objects which invoke hotplug 11230 * without deadlocking with linkwatch via keventd. 11231 * 2) Since we run with the RTNL semaphore not held, we can sleep 11232 * safely in order to wait for the netdev refcnt to drop to zero. 11233 * 11234 * We must not return until all unregister events added during 11235 * the interval the lock was held have been completed. 11236 */ 11237 void netdev_run_todo(void) 11238 { 11239 struct net_device *dev, *tmp; 11240 struct list_head list; 11241 int cnt; 11242 #ifdef CONFIG_LOCKDEP 11243 struct list_head unlink_list; 11244 11245 list_replace_init(&net_unlink_list, &unlink_list); 11246 11247 while (!list_empty(&unlink_list)) { 11248 dev = list_first_entry(&unlink_list, struct net_device, 11249 unlink_list); 11250 list_del_init(&dev->unlink_list); 11251 dev->nested_level = dev->lower_level - 1; 11252 } 11253 #endif 11254 11255 /* Snapshot list, allow later requests */ 11256 list_replace_init(&net_todo_list, &list); 11257 11258 __rtnl_unlock(); 11259 11260 /* Wait for rcu callbacks to finish before next phase */ 11261 if (!list_empty(&list)) 11262 rcu_barrier(); 11263 11264 list_for_each_entry_safe(dev, tmp, &list, todo_list) { 11265 if (unlikely(dev->reg_state != NETREG_UNREGISTERING)) { 11266 netdev_WARN(dev, "run_todo but not unregistering\n"); 11267 list_del(&dev->todo_list); 11268 continue; 11269 } 11270 11271 netdev_lock(dev); 11272 WRITE_ONCE(dev->reg_state, NETREG_UNREGISTERED); 11273 netdev_unlock(dev); 11274 linkwatch_sync_dev(dev); 11275 } 11276 11277 cnt = 0; 11278 while (!list_empty(&list)) { 11279 dev = netdev_wait_allrefs_any(&list); 11280 list_del(&dev->todo_list); 11281 11282 /* paranoia */ 11283 BUG_ON(netdev_refcnt_read(dev) != 1); 11284 BUG_ON(!list_empty(&dev->ptype_all)); 11285 BUG_ON(!list_empty(&dev->ptype_specific)); 11286 WARN_ON(rcu_access_pointer(dev->ip_ptr)); 11287 WARN_ON(rcu_access_pointer(dev->ip6_ptr)); 11288 11289 netdev_do_free_pcpu_stats(dev); 11290 if (dev->priv_destructor) 11291 dev->priv_destructor(dev); 11292 if (dev->needs_free_netdev) 11293 free_netdev(dev); 11294 11295 cnt++; 11296 11297 /* Free network device */ 11298 kobject_put(&dev->dev.kobj); 11299 } 11300 if (cnt && atomic_sub_and_test(cnt, &dev_unreg_count)) 11301 wake_up(&netdev_unregistering_wq); 11302 } 11303 11304 /* Collate per-cpu network dstats statistics 11305 * 11306 * Read per-cpu network statistics from dev->dstats and populate the related 11307 * fields in @s. 11308 */ 11309 static void dev_fetch_dstats(struct rtnl_link_stats64 *s, 11310 const struct pcpu_dstats __percpu *dstats) 11311 { 11312 int cpu; 11313 11314 for_each_possible_cpu(cpu) { 11315 u64 rx_packets, rx_bytes, rx_drops; 11316 u64 tx_packets, tx_bytes, tx_drops; 11317 const struct pcpu_dstats *stats; 11318 unsigned int start; 11319 11320 stats = per_cpu_ptr(dstats, cpu); 11321 do { 11322 start = u64_stats_fetch_begin(&stats->syncp); 11323 rx_packets = u64_stats_read(&stats->rx_packets); 11324 rx_bytes = u64_stats_read(&stats->rx_bytes); 11325 rx_drops = u64_stats_read(&stats->rx_drops); 11326 tx_packets = u64_stats_read(&stats->tx_packets); 11327 tx_bytes = u64_stats_read(&stats->tx_bytes); 11328 tx_drops = u64_stats_read(&stats->tx_drops); 11329 } while (u64_stats_fetch_retry(&stats->syncp, start)); 11330 11331 s->rx_packets += rx_packets; 11332 s->rx_bytes += rx_bytes; 11333 s->rx_dropped += rx_drops; 11334 s->tx_packets += tx_packets; 11335 s->tx_bytes += tx_bytes; 11336 s->tx_dropped += tx_drops; 11337 } 11338 } 11339 11340 /* ndo_get_stats64 implementation for dtstats-based accounting. 11341 * 11342 * Populate @s from dev->stats and dev->dstats. This is used internally by the 11343 * core for NETDEV_PCPU_STAT_DSTAT-type stats collection. 11344 */ 11345 static void dev_get_dstats64(const struct net_device *dev, 11346 struct rtnl_link_stats64 *s) 11347 { 11348 netdev_stats_to_stats64(s, &dev->stats); 11349 dev_fetch_dstats(s, dev->dstats); 11350 } 11351 11352 /* Convert net_device_stats to rtnl_link_stats64. rtnl_link_stats64 has 11353 * all the same fields in the same order as net_device_stats, with only 11354 * the type differing, but rtnl_link_stats64 may have additional fields 11355 * at the end for newer counters. 11356 */ 11357 void netdev_stats_to_stats64(struct rtnl_link_stats64 *stats64, 11358 const struct net_device_stats *netdev_stats) 11359 { 11360 size_t i, n = sizeof(*netdev_stats) / sizeof(atomic_long_t); 11361 const atomic_long_t *src = (atomic_long_t *)netdev_stats; 11362 u64 *dst = (u64 *)stats64; 11363 11364 BUILD_BUG_ON(n > sizeof(*stats64) / sizeof(u64)); 11365 for (i = 0; i < n; i++) 11366 dst[i] = (unsigned long)atomic_long_read(&src[i]); 11367 /* zero out counters that only exist in rtnl_link_stats64 */ 11368 memset((char *)stats64 + n * sizeof(u64), 0, 11369 sizeof(*stats64) - n * sizeof(u64)); 11370 } 11371 EXPORT_SYMBOL(netdev_stats_to_stats64); 11372 11373 static __cold struct net_device_core_stats __percpu *netdev_core_stats_alloc( 11374 struct net_device *dev) 11375 { 11376 struct net_device_core_stats __percpu *p; 11377 11378 p = alloc_percpu_gfp(struct net_device_core_stats, 11379 GFP_ATOMIC | __GFP_NOWARN); 11380 11381 if (p && cmpxchg(&dev->core_stats, NULL, p)) 11382 free_percpu(p); 11383 11384 /* This READ_ONCE() pairs with the cmpxchg() above */ 11385 return READ_ONCE(dev->core_stats); 11386 } 11387 11388 noinline void netdev_core_stats_inc(struct net_device *dev, u32 offset) 11389 { 11390 /* This READ_ONCE() pairs with the write in netdev_core_stats_alloc() */ 11391 struct net_device_core_stats __percpu *p = READ_ONCE(dev->core_stats); 11392 unsigned long __percpu *field; 11393 11394 if (unlikely(!p)) { 11395 p = netdev_core_stats_alloc(dev); 11396 if (!p) 11397 return; 11398 } 11399 11400 field = (unsigned long __percpu *)((void __percpu *)p + offset); 11401 this_cpu_inc(*field); 11402 } 11403 EXPORT_SYMBOL_GPL(netdev_core_stats_inc); 11404 11405 /** 11406 * dev_get_stats - get network device statistics 11407 * @dev: device to get statistics from 11408 * @storage: place to store stats 11409 * 11410 * Get network statistics from device. Return @storage. 11411 * The device driver may provide its own method by setting 11412 * dev->netdev_ops->get_stats64 or dev->netdev_ops->get_stats; 11413 * otherwise the internal statistics structure is used. 11414 */ 11415 struct rtnl_link_stats64 *dev_get_stats(struct net_device *dev, 11416 struct rtnl_link_stats64 *storage) 11417 { 11418 const struct net_device_ops *ops = dev->netdev_ops; 11419 const struct net_device_core_stats __percpu *p; 11420 11421 /* 11422 * IPv{4,6} and udp tunnels share common stat helpers and use 11423 * different stat type (NETDEV_PCPU_STAT_TSTATS vs 11424 * NETDEV_PCPU_STAT_DSTATS). Ensure the accounting is consistent. 11425 */ 11426 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, rx_bytes) != 11427 offsetof(struct pcpu_dstats, rx_bytes)); 11428 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, rx_packets) != 11429 offsetof(struct pcpu_dstats, rx_packets)); 11430 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, tx_bytes) != 11431 offsetof(struct pcpu_dstats, tx_bytes)); 11432 BUILD_BUG_ON(offsetof(struct pcpu_sw_netstats, tx_packets) != 11433 offsetof(struct pcpu_dstats, tx_packets)); 11434 11435 if (ops->ndo_get_stats64) { 11436 memset(storage, 0, sizeof(*storage)); 11437 ops->ndo_get_stats64(dev, storage); 11438 } else if (ops->ndo_get_stats) { 11439 netdev_stats_to_stats64(storage, ops->ndo_get_stats(dev)); 11440 } else if (dev->pcpu_stat_type == NETDEV_PCPU_STAT_TSTATS) { 11441 dev_get_tstats64(dev, storage); 11442 } else if (dev->pcpu_stat_type == NETDEV_PCPU_STAT_DSTATS) { 11443 dev_get_dstats64(dev, storage); 11444 } else { 11445 netdev_stats_to_stats64(storage, &dev->stats); 11446 } 11447 11448 /* This READ_ONCE() pairs with the write in netdev_core_stats_alloc() */ 11449 p = READ_ONCE(dev->core_stats); 11450 if (p) { 11451 const struct net_device_core_stats *core_stats; 11452 int i; 11453 11454 for_each_possible_cpu(i) { 11455 core_stats = per_cpu_ptr(p, i); 11456 storage->rx_dropped += READ_ONCE(core_stats->rx_dropped); 11457 storage->tx_dropped += READ_ONCE(core_stats->tx_dropped); 11458 storage->rx_nohandler += READ_ONCE(core_stats->rx_nohandler); 11459 storage->rx_otherhost_dropped += READ_ONCE(core_stats->rx_otherhost_dropped); 11460 } 11461 } 11462 return storage; 11463 } 11464 EXPORT_SYMBOL(dev_get_stats); 11465 11466 /** 11467 * dev_fetch_sw_netstats - get per-cpu network device statistics 11468 * @s: place to store stats 11469 * @netstats: per-cpu network stats to read from 11470 * 11471 * Read per-cpu network statistics and populate the related fields in @s. 11472 */ 11473 void dev_fetch_sw_netstats(struct rtnl_link_stats64 *s, 11474 const struct pcpu_sw_netstats __percpu *netstats) 11475 { 11476 int cpu; 11477 11478 for_each_possible_cpu(cpu) { 11479 u64 rx_packets, rx_bytes, tx_packets, tx_bytes; 11480 const struct pcpu_sw_netstats *stats; 11481 unsigned int start; 11482 11483 stats = per_cpu_ptr(netstats, cpu); 11484 do { 11485 start = u64_stats_fetch_begin(&stats->syncp); 11486 rx_packets = u64_stats_read(&stats->rx_packets); 11487 rx_bytes = u64_stats_read(&stats->rx_bytes); 11488 tx_packets = u64_stats_read(&stats->tx_packets); 11489 tx_bytes = u64_stats_read(&stats->tx_bytes); 11490 } while (u64_stats_fetch_retry(&stats->syncp, start)); 11491 11492 s->rx_packets += rx_packets; 11493 s->rx_bytes += rx_bytes; 11494 s->tx_packets += tx_packets; 11495 s->tx_bytes += tx_bytes; 11496 } 11497 } 11498 EXPORT_SYMBOL_GPL(dev_fetch_sw_netstats); 11499 11500 /** 11501 * dev_get_tstats64 - ndo_get_stats64 implementation 11502 * @dev: device to get statistics from 11503 * @s: place to store stats 11504 * 11505 * Populate @s from dev->stats and dev->tstats. Can be used as 11506 * ndo_get_stats64() callback. 11507 */ 11508 void dev_get_tstats64(struct net_device *dev, struct rtnl_link_stats64 *s) 11509 { 11510 netdev_stats_to_stats64(s, &dev->stats); 11511 dev_fetch_sw_netstats(s, dev->tstats); 11512 } 11513 EXPORT_SYMBOL_GPL(dev_get_tstats64); 11514 11515 struct netdev_queue *dev_ingress_queue_create(struct net_device *dev) 11516 { 11517 struct netdev_queue *queue = dev_ingress_queue(dev); 11518 11519 #ifdef CONFIG_NET_CLS_ACT 11520 if (queue) 11521 return queue; 11522 queue = kzalloc(sizeof(*queue), GFP_KERNEL); 11523 if (!queue) 11524 return NULL; 11525 netdev_init_one_queue(dev, queue, NULL); 11526 RCU_INIT_POINTER(queue->qdisc, &noop_qdisc); 11527 RCU_INIT_POINTER(queue->qdisc_sleeping, &noop_qdisc); 11528 rcu_assign_pointer(dev->ingress_queue, queue); 11529 #endif 11530 return queue; 11531 } 11532 11533 static const struct ethtool_ops default_ethtool_ops; 11534 11535 void netdev_set_default_ethtool_ops(struct net_device *dev, 11536 const struct ethtool_ops *ops) 11537 { 11538 if (dev->ethtool_ops == &default_ethtool_ops) 11539 dev->ethtool_ops = ops; 11540 } 11541 EXPORT_SYMBOL_GPL(netdev_set_default_ethtool_ops); 11542 11543 /** 11544 * netdev_sw_irq_coalesce_default_on() - enable SW IRQ coalescing by default 11545 * @dev: netdev to enable the IRQ coalescing on 11546 * 11547 * Sets a conservative default for SW IRQ coalescing. Users can use 11548 * sysfs attributes to override the default values. 11549 */ 11550 void netdev_sw_irq_coalesce_default_on(struct net_device *dev) 11551 { 11552 WARN_ON(dev->reg_state == NETREG_REGISTERED); 11553 11554 if (!IS_ENABLED(CONFIG_PREEMPT_RT)) { 11555 netdev_set_gro_flush_timeout(dev, 20000); 11556 netdev_set_defer_hard_irqs(dev, 1); 11557 } 11558 } 11559 EXPORT_SYMBOL_GPL(netdev_sw_irq_coalesce_default_on); 11560 11561 /** 11562 * alloc_netdev_mqs - allocate network device 11563 * @sizeof_priv: size of private data to allocate space for 11564 * @name: device name format string 11565 * @name_assign_type: origin of device name 11566 * @setup: callback to initialize device 11567 * @txqs: the number of TX subqueues to allocate 11568 * @rxqs: the number of RX subqueues to allocate 11569 * 11570 * Allocates a struct net_device with private data area for driver use 11571 * and performs basic initialization. Also allocates subqueue structs 11572 * for each queue on the device. 11573 */ 11574 struct net_device *alloc_netdev_mqs(int sizeof_priv, const char *name, 11575 unsigned char name_assign_type, 11576 void (*setup)(struct net_device *), 11577 unsigned int txqs, unsigned int rxqs) 11578 { 11579 struct net_device *dev; 11580 size_t napi_config_sz; 11581 unsigned int maxqs; 11582 11583 BUG_ON(strlen(name) >= sizeof(dev->name)); 11584 11585 if (txqs < 1) { 11586 pr_err("alloc_netdev: Unable to allocate device with zero queues\n"); 11587 return NULL; 11588 } 11589 11590 if (rxqs < 1) { 11591 pr_err("alloc_netdev: Unable to allocate device with zero RX queues\n"); 11592 return NULL; 11593 } 11594 11595 maxqs = max(txqs, rxqs); 11596 11597 dev = kvzalloc(struct_size(dev, priv, sizeof_priv), 11598 GFP_KERNEL_ACCOUNT | __GFP_RETRY_MAYFAIL); 11599 if (!dev) 11600 return NULL; 11601 11602 dev->priv_len = sizeof_priv; 11603 11604 ref_tracker_dir_init(&dev->refcnt_tracker, 128, name); 11605 #ifdef CONFIG_PCPU_DEV_REFCNT 11606 dev->pcpu_refcnt = alloc_percpu(int); 11607 if (!dev->pcpu_refcnt) 11608 goto free_dev; 11609 __dev_hold(dev); 11610 #else 11611 refcount_set(&dev->dev_refcnt, 1); 11612 #endif 11613 11614 if (dev_addr_init(dev)) 11615 goto free_pcpu; 11616 11617 dev_mc_init(dev); 11618 dev_uc_init(dev); 11619 11620 dev_net_set(dev, &init_net); 11621 11622 dev->gso_max_size = GSO_LEGACY_MAX_SIZE; 11623 dev->xdp_zc_max_segs = 1; 11624 dev->gso_max_segs = GSO_MAX_SEGS; 11625 dev->gro_max_size = GRO_LEGACY_MAX_SIZE; 11626 dev->gso_ipv4_max_size = GSO_LEGACY_MAX_SIZE; 11627 dev->gro_ipv4_max_size = GRO_LEGACY_MAX_SIZE; 11628 dev->tso_max_size = TSO_LEGACY_MAX_SIZE; 11629 dev->tso_max_segs = TSO_MAX_SEGS; 11630 dev->upper_level = 1; 11631 dev->lower_level = 1; 11632 #ifdef CONFIG_LOCKDEP 11633 dev->nested_level = 0; 11634 INIT_LIST_HEAD(&dev->unlink_list); 11635 #endif 11636 11637 INIT_LIST_HEAD(&dev->napi_list); 11638 INIT_LIST_HEAD(&dev->unreg_list); 11639 INIT_LIST_HEAD(&dev->close_list); 11640 INIT_LIST_HEAD(&dev->link_watch_list); 11641 INIT_LIST_HEAD(&dev->adj_list.upper); 11642 INIT_LIST_HEAD(&dev->adj_list.lower); 11643 INIT_LIST_HEAD(&dev->ptype_all); 11644 INIT_LIST_HEAD(&dev->ptype_specific); 11645 INIT_LIST_HEAD(&dev->net_notifier_list); 11646 #ifdef CONFIG_NET_SCHED 11647 hash_init(dev->qdisc_hash); 11648 #endif 11649 11650 mutex_init(&dev->lock); 11651 11652 dev->priv_flags = IFF_XMIT_DST_RELEASE | IFF_XMIT_DST_RELEASE_PERM; 11653 setup(dev); 11654 11655 if (!dev->tx_queue_len) { 11656 dev->priv_flags |= IFF_NO_QUEUE; 11657 dev->tx_queue_len = DEFAULT_TX_QUEUE_LEN; 11658 } 11659 11660 dev->num_tx_queues = txqs; 11661 dev->real_num_tx_queues = txqs; 11662 if (netif_alloc_netdev_queues(dev)) 11663 goto free_all; 11664 11665 dev->num_rx_queues = rxqs; 11666 dev->real_num_rx_queues = rxqs; 11667 if (netif_alloc_rx_queues(dev)) 11668 goto free_all; 11669 dev->ethtool = kzalloc(sizeof(*dev->ethtool), GFP_KERNEL_ACCOUNT); 11670 if (!dev->ethtool) 11671 goto free_all; 11672 11673 dev->cfg = kzalloc(sizeof(*dev->cfg), GFP_KERNEL_ACCOUNT); 11674 if (!dev->cfg) 11675 goto free_all; 11676 dev->cfg_pending = dev->cfg; 11677 11678 napi_config_sz = array_size(maxqs, sizeof(*dev->napi_config)); 11679 dev->napi_config = kvzalloc(napi_config_sz, GFP_KERNEL_ACCOUNT); 11680 if (!dev->napi_config) 11681 goto free_all; 11682 11683 strscpy(dev->name, name); 11684 dev->name_assign_type = name_assign_type; 11685 dev->group = INIT_NETDEV_GROUP; 11686 if (!dev->ethtool_ops) 11687 dev->ethtool_ops = &default_ethtool_ops; 11688 11689 nf_hook_netdev_init(dev); 11690 11691 return dev; 11692 11693 free_all: 11694 free_netdev(dev); 11695 return NULL; 11696 11697 free_pcpu: 11698 #ifdef CONFIG_PCPU_DEV_REFCNT 11699 free_percpu(dev->pcpu_refcnt); 11700 free_dev: 11701 #endif 11702 kvfree(dev); 11703 return NULL; 11704 } 11705 EXPORT_SYMBOL(alloc_netdev_mqs); 11706 11707 static void netdev_napi_exit(struct net_device *dev) 11708 { 11709 if (!list_empty(&dev->napi_list)) { 11710 struct napi_struct *p, *n; 11711 11712 netdev_lock(dev); 11713 list_for_each_entry_safe(p, n, &dev->napi_list, dev_list) 11714 __netif_napi_del_locked(p); 11715 netdev_unlock(dev); 11716 11717 synchronize_net(); 11718 } 11719 11720 kvfree(dev->napi_config); 11721 } 11722 11723 /** 11724 * free_netdev - free network device 11725 * @dev: device 11726 * 11727 * This function does the last stage of destroying an allocated device 11728 * interface. The reference to the device object is released. If this 11729 * is the last reference then it will be freed.Must be called in process 11730 * context. 11731 */ 11732 void free_netdev(struct net_device *dev) 11733 { 11734 might_sleep(); 11735 11736 /* When called immediately after register_netdevice() failed the unwind 11737 * handling may still be dismantling the device. Handle that case by 11738 * deferring the free. 11739 */ 11740 if (dev->reg_state == NETREG_UNREGISTERING) { 11741 ASSERT_RTNL(); 11742 dev->needs_free_netdev = true; 11743 return; 11744 } 11745 11746 WARN_ON(dev->cfg != dev->cfg_pending); 11747 kfree(dev->cfg); 11748 kfree(dev->ethtool); 11749 netif_free_tx_queues(dev); 11750 netif_free_rx_queues(dev); 11751 11752 kfree(rcu_dereference_protected(dev->ingress_queue, 1)); 11753 11754 /* Flush device addresses */ 11755 dev_addr_flush(dev); 11756 11757 netdev_napi_exit(dev); 11758 11759 netif_del_cpu_rmap(dev); 11760 11761 ref_tracker_dir_exit(&dev->refcnt_tracker); 11762 #ifdef CONFIG_PCPU_DEV_REFCNT 11763 free_percpu(dev->pcpu_refcnt); 11764 dev->pcpu_refcnt = NULL; 11765 #endif 11766 free_percpu(dev->core_stats); 11767 dev->core_stats = NULL; 11768 free_percpu(dev->xdp_bulkq); 11769 dev->xdp_bulkq = NULL; 11770 11771 netdev_free_phy_link_topology(dev); 11772 11773 mutex_destroy(&dev->lock); 11774 11775 /* Compatibility with error handling in drivers */ 11776 if (dev->reg_state == NETREG_UNINITIALIZED || 11777 dev->reg_state == NETREG_DUMMY) { 11778 kvfree(dev); 11779 return; 11780 } 11781 11782 BUG_ON(dev->reg_state != NETREG_UNREGISTERED); 11783 WRITE_ONCE(dev->reg_state, NETREG_RELEASED); 11784 11785 /* will free via device release */ 11786 put_device(&dev->dev); 11787 } 11788 EXPORT_SYMBOL(free_netdev); 11789 11790 /** 11791 * alloc_netdev_dummy - Allocate and initialize a dummy net device. 11792 * @sizeof_priv: size of private data to allocate space for 11793 * 11794 * Return: the allocated net_device on success, NULL otherwise 11795 */ 11796 struct net_device *alloc_netdev_dummy(int sizeof_priv) 11797 { 11798 return alloc_netdev(sizeof_priv, "dummy#", NET_NAME_UNKNOWN, 11799 init_dummy_netdev); 11800 } 11801 EXPORT_SYMBOL_GPL(alloc_netdev_dummy); 11802 11803 /** 11804 * synchronize_net - Synchronize with packet receive processing 11805 * 11806 * Wait for packets currently being received to be done. 11807 * Does not block later packets from starting. 11808 */ 11809 void synchronize_net(void) 11810 { 11811 might_sleep(); 11812 if (from_cleanup_net() || rtnl_is_locked()) 11813 synchronize_rcu_expedited(); 11814 else 11815 synchronize_rcu(); 11816 } 11817 EXPORT_SYMBOL(synchronize_net); 11818 11819 static void netdev_rss_contexts_free(struct net_device *dev) 11820 { 11821 struct ethtool_rxfh_context *ctx; 11822 unsigned long context; 11823 11824 mutex_lock(&dev->ethtool->rss_lock); 11825 xa_for_each(&dev->ethtool->rss_ctx, context, ctx) { 11826 struct ethtool_rxfh_param rxfh; 11827 11828 rxfh.indir = ethtool_rxfh_context_indir(ctx); 11829 rxfh.key = ethtool_rxfh_context_key(ctx); 11830 rxfh.hfunc = ctx->hfunc; 11831 rxfh.input_xfrm = ctx->input_xfrm; 11832 rxfh.rss_context = context; 11833 rxfh.rss_delete = true; 11834 11835 xa_erase(&dev->ethtool->rss_ctx, context); 11836 if (dev->ethtool_ops->create_rxfh_context) 11837 dev->ethtool_ops->remove_rxfh_context(dev, ctx, 11838 context, NULL); 11839 else 11840 dev->ethtool_ops->set_rxfh(dev, &rxfh, NULL); 11841 kfree(ctx); 11842 } 11843 xa_destroy(&dev->ethtool->rss_ctx); 11844 mutex_unlock(&dev->ethtool->rss_lock); 11845 } 11846 11847 /** 11848 * unregister_netdevice_queue - remove device from the kernel 11849 * @dev: device 11850 * @head: list 11851 * 11852 * This function shuts down a device interface and removes it 11853 * from the kernel tables. 11854 * If head not NULL, device is queued to be unregistered later. 11855 * 11856 * Callers must hold the rtnl semaphore. You may want 11857 * unregister_netdev() instead of this. 11858 */ 11859 11860 void unregister_netdevice_queue(struct net_device *dev, struct list_head *head) 11861 { 11862 ASSERT_RTNL(); 11863 11864 if (head) { 11865 list_move_tail(&dev->unreg_list, head); 11866 } else { 11867 LIST_HEAD(single); 11868 11869 list_add(&dev->unreg_list, &single); 11870 unregister_netdevice_many(&single); 11871 } 11872 } 11873 EXPORT_SYMBOL(unregister_netdevice_queue); 11874 11875 static void dev_memory_provider_uninstall(struct net_device *dev) 11876 { 11877 unsigned int i; 11878 11879 for (i = 0; i < dev->real_num_rx_queues; i++) { 11880 struct netdev_rx_queue *rxq = &dev->_rx[i]; 11881 struct pp_memory_provider_params *p = &rxq->mp_params; 11882 11883 if (p->mp_ops && p->mp_ops->uninstall) 11884 p->mp_ops->uninstall(rxq->mp_params.mp_priv, rxq); 11885 } 11886 } 11887 11888 void unregister_netdevice_many_notify(struct list_head *head, 11889 u32 portid, const struct nlmsghdr *nlh) 11890 { 11891 struct net_device *dev, *tmp; 11892 LIST_HEAD(close_head); 11893 int cnt = 0; 11894 11895 BUG_ON(dev_boot_phase); 11896 ASSERT_RTNL(); 11897 11898 if (list_empty(head)) 11899 return; 11900 11901 list_for_each_entry_safe(dev, tmp, head, unreg_list) { 11902 /* Some devices call without registering 11903 * for initialization unwind. Remove those 11904 * devices and proceed with the remaining. 11905 */ 11906 if (dev->reg_state == NETREG_UNINITIALIZED) { 11907 pr_debug("unregister_netdevice: device %s/%p never was registered\n", 11908 dev->name, dev); 11909 11910 WARN_ON(1); 11911 list_del(&dev->unreg_list); 11912 continue; 11913 } 11914 dev->dismantle = true; 11915 BUG_ON(dev->reg_state != NETREG_REGISTERED); 11916 } 11917 11918 /* If device is running, close it first. */ 11919 list_for_each_entry(dev, head, unreg_list) { 11920 list_add_tail(&dev->close_list, &close_head); 11921 netdev_lock_ops(dev); 11922 } 11923 dev_close_many(&close_head, true); 11924 11925 list_for_each_entry(dev, head, unreg_list) { 11926 netdev_unlock_ops(dev); 11927 /* And unlink it from device chain. */ 11928 unlist_netdevice(dev); 11929 netdev_lock(dev); 11930 WRITE_ONCE(dev->reg_state, NETREG_UNREGISTERING); 11931 netdev_unlock(dev); 11932 } 11933 flush_all_backlogs(); 11934 11935 synchronize_net(); 11936 11937 list_for_each_entry(dev, head, unreg_list) { 11938 struct sk_buff *skb = NULL; 11939 11940 /* Shutdown queueing discipline. */ 11941 dev_shutdown(dev); 11942 dev_tcx_uninstall(dev); 11943 netdev_lock_ops(dev); 11944 dev_xdp_uninstall(dev); 11945 netdev_unlock_ops(dev); 11946 bpf_dev_bound_netdev_unregister(dev); 11947 dev_memory_provider_uninstall(dev); 11948 11949 netdev_offload_xstats_disable_all(dev); 11950 11951 /* Notify protocols, that we are about to destroy 11952 * this device. They should clean all the things. 11953 */ 11954 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 11955 11956 if (!dev->rtnl_link_ops || 11957 dev->rtnl_link_state == RTNL_LINK_INITIALIZED) 11958 skb = rtmsg_ifinfo_build_skb(RTM_DELLINK, dev, ~0U, 0, 11959 GFP_KERNEL, NULL, 0, 11960 portid, nlh); 11961 11962 /* 11963 * Flush the unicast and multicast chains 11964 */ 11965 dev_uc_flush(dev); 11966 dev_mc_flush(dev); 11967 11968 netdev_name_node_alt_flush(dev); 11969 netdev_name_node_free(dev->name_node); 11970 11971 netdev_rss_contexts_free(dev); 11972 11973 call_netdevice_notifiers(NETDEV_PRE_UNINIT, dev); 11974 11975 if (dev->netdev_ops->ndo_uninit) 11976 dev->netdev_ops->ndo_uninit(dev); 11977 11978 mutex_destroy(&dev->ethtool->rss_lock); 11979 11980 net_shaper_flush_netdev(dev); 11981 11982 if (skb) 11983 rtmsg_ifinfo_send(skb, dev, GFP_KERNEL, portid, nlh); 11984 11985 /* Notifier chain MUST detach us all upper devices. */ 11986 WARN_ON(netdev_has_any_upper_dev(dev)); 11987 WARN_ON(netdev_has_any_lower_dev(dev)); 11988 11989 /* Remove entries from kobject tree */ 11990 netdev_unregister_kobject(dev); 11991 #ifdef CONFIG_XPS 11992 /* Remove XPS queueing entries */ 11993 netif_reset_xps_queues_gt(dev, 0); 11994 #endif 11995 } 11996 11997 synchronize_net(); 11998 11999 list_for_each_entry(dev, head, unreg_list) { 12000 netdev_put(dev, &dev->dev_registered_tracker); 12001 net_set_todo(dev); 12002 cnt++; 12003 } 12004 atomic_add(cnt, &dev_unreg_count); 12005 12006 list_del(head); 12007 } 12008 12009 /** 12010 * unregister_netdevice_many - unregister many devices 12011 * @head: list of devices 12012 * 12013 * Note: As most callers use a stack allocated list_head, 12014 * we force a list_del() to make sure stack won't be corrupted later. 12015 */ 12016 void unregister_netdevice_many(struct list_head *head) 12017 { 12018 unregister_netdevice_many_notify(head, 0, NULL); 12019 } 12020 EXPORT_SYMBOL(unregister_netdevice_many); 12021 12022 /** 12023 * unregister_netdev - remove device from the kernel 12024 * @dev: device 12025 * 12026 * This function shuts down a device interface and removes it 12027 * from the kernel tables. 12028 * 12029 * This is just a wrapper for unregister_netdevice that takes 12030 * the rtnl semaphore. In general you want to use this and not 12031 * unregister_netdevice. 12032 */ 12033 void unregister_netdev(struct net_device *dev) 12034 { 12035 rtnl_net_dev_lock(dev); 12036 unregister_netdevice(dev); 12037 rtnl_net_dev_unlock(dev); 12038 } 12039 EXPORT_SYMBOL(unregister_netdev); 12040 12041 int netif_change_net_namespace(struct net_device *dev, struct net *net, 12042 const char *pat, int new_ifindex, 12043 struct netlink_ext_ack *extack) 12044 { 12045 struct netdev_name_node *name_node; 12046 struct net *net_old = dev_net(dev); 12047 char new_name[IFNAMSIZ] = {}; 12048 int err, new_nsid; 12049 12050 ASSERT_RTNL(); 12051 12052 /* Don't allow namespace local devices to be moved. */ 12053 err = -EINVAL; 12054 if (dev->netns_immutable) { 12055 NL_SET_ERR_MSG(extack, "The interface netns is immutable"); 12056 goto out; 12057 } 12058 12059 /* Ensure the device has been registered */ 12060 if (dev->reg_state != NETREG_REGISTERED) { 12061 NL_SET_ERR_MSG(extack, "The interface isn't registered"); 12062 goto out; 12063 } 12064 12065 /* Get out if there is nothing todo */ 12066 err = 0; 12067 if (net_eq(net_old, net)) 12068 goto out; 12069 12070 /* Pick the destination device name, and ensure 12071 * we can use it in the destination network namespace. 12072 */ 12073 err = -EEXIST; 12074 if (netdev_name_in_use(net, dev->name)) { 12075 /* We get here if we can't use the current device name */ 12076 if (!pat) { 12077 NL_SET_ERR_MSG(extack, 12078 "An interface with the same name exists in the target netns"); 12079 goto out; 12080 } 12081 err = dev_prep_valid_name(net, dev, pat, new_name, EEXIST); 12082 if (err < 0) { 12083 NL_SET_ERR_MSG_FMT(extack, 12084 "Unable to use '%s' for the new interface name in the target netns", 12085 pat); 12086 goto out; 12087 } 12088 } 12089 /* Check that none of the altnames conflicts. */ 12090 err = -EEXIST; 12091 netdev_for_each_altname(dev, name_node) { 12092 if (netdev_name_in_use(net, name_node->name)) { 12093 NL_SET_ERR_MSG_FMT(extack, 12094 "An interface with the altname %s exists in the target netns", 12095 name_node->name); 12096 goto out; 12097 } 12098 } 12099 12100 /* Check that new_ifindex isn't used yet. */ 12101 if (new_ifindex) { 12102 err = dev_index_reserve(net, new_ifindex); 12103 if (err < 0) { 12104 NL_SET_ERR_MSG_FMT(extack, 12105 "The ifindex %d is not available in the target netns", 12106 new_ifindex); 12107 goto out; 12108 } 12109 } else { 12110 /* If there is an ifindex conflict assign a new one */ 12111 err = dev_index_reserve(net, dev->ifindex); 12112 if (err == -EBUSY) 12113 err = dev_index_reserve(net, 0); 12114 if (err < 0) { 12115 NL_SET_ERR_MSG(extack, 12116 "Unable to allocate a new ifindex in the target netns"); 12117 goto out; 12118 } 12119 new_ifindex = err; 12120 } 12121 12122 /* 12123 * And now a mini version of register_netdevice unregister_netdevice. 12124 */ 12125 12126 /* If device is running close it first. */ 12127 netif_close(dev); 12128 12129 /* And unlink it from device chain */ 12130 unlist_netdevice(dev); 12131 12132 synchronize_net(); 12133 12134 /* Shutdown queueing discipline. */ 12135 dev_shutdown(dev); 12136 12137 /* Notify protocols, that we are about to destroy 12138 * this device. They should clean all the things. 12139 * 12140 * Note that dev->reg_state stays at NETREG_REGISTERED. 12141 * This is wanted because this way 8021q and macvlan know 12142 * the device is just moving and can keep their slaves up. 12143 */ 12144 call_netdevice_notifiers(NETDEV_UNREGISTER, dev); 12145 rcu_barrier(); 12146 12147 new_nsid = peernet2id_alloc(dev_net(dev), net, GFP_KERNEL); 12148 12149 rtmsg_ifinfo_newnet(RTM_DELLINK, dev, ~0U, GFP_KERNEL, &new_nsid, 12150 new_ifindex); 12151 12152 /* 12153 * Flush the unicast and multicast chains 12154 */ 12155 dev_uc_flush(dev); 12156 dev_mc_flush(dev); 12157 12158 /* Send a netdev-removed uevent to the old namespace */ 12159 kobject_uevent(&dev->dev.kobj, KOBJ_REMOVE); 12160 netdev_adjacent_del_links(dev); 12161 12162 /* Move per-net netdevice notifiers that are following the netdevice */ 12163 move_netdevice_notifiers_dev_net(dev, net); 12164 12165 /* Actually switch the network namespace */ 12166 dev_net_set(dev, net); 12167 dev->ifindex = new_ifindex; 12168 12169 if (new_name[0]) { 12170 /* Rename the netdev to prepared name */ 12171 write_seqlock_bh(&netdev_rename_lock); 12172 strscpy(dev->name, new_name, IFNAMSIZ); 12173 write_sequnlock_bh(&netdev_rename_lock); 12174 } 12175 12176 /* Fixup kobjects */ 12177 dev_set_uevent_suppress(&dev->dev, 1); 12178 err = device_rename(&dev->dev, dev->name); 12179 dev_set_uevent_suppress(&dev->dev, 0); 12180 WARN_ON(err); 12181 12182 /* Send a netdev-add uevent to the new namespace */ 12183 kobject_uevent(&dev->dev.kobj, KOBJ_ADD); 12184 netdev_adjacent_add_links(dev); 12185 12186 /* Adapt owner in case owning user namespace of target network 12187 * namespace is different from the original one. 12188 */ 12189 err = netdev_change_owner(dev, net_old, net); 12190 WARN_ON(err); 12191 12192 /* Add the device back in the hashes */ 12193 list_netdevice(dev); 12194 12195 /* Notify protocols, that a new device appeared. */ 12196 call_netdevice_notifiers(NETDEV_REGISTER, dev); 12197 12198 /* 12199 * Prevent userspace races by waiting until the network 12200 * device is fully setup before sending notifications. 12201 */ 12202 rtmsg_ifinfo(RTM_NEWLINK, dev, ~0U, GFP_KERNEL, 0, NULL); 12203 12204 synchronize_net(); 12205 err = 0; 12206 out: 12207 return err; 12208 } 12209 12210 static int dev_cpu_dead(unsigned int oldcpu) 12211 { 12212 struct sk_buff **list_skb; 12213 struct sk_buff *skb; 12214 unsigned int cpu; 12215 struct softnet_data *sd, *oldsd, *remsd = NULL; 12216 12217 local_irq_disable(); 12218 cpu = smp_processor_id(); 12219 sd = &per_cpu(softnet_data, cpu); 12220 oldsd = &per_cpu(softnet_data, oldcpu); 12221 12222 /* Find end of our completion_queue. */ 12223 list_skb = &sd->completion_queue; 12224 while (*list_skb) 12225 list_skb = &(*list_skb)->next; 12226 /* Append completion queue from offline CPU. */ 12227 *list_skb = oldsd->completion_queue; 12228 oldsd->completion_queue = NULL; 12229 12230 /* Append output queue from offline CPU. */ 12231 if (oldsd->output_queue) { 12232 *sd->output_queue_tailp = oldsd->output_queue; 12233 sd->output_queue_tailp = oldsd->output_queue_tailp; 12234 oldsd->output_queue = NULL; 12235 oldsd->output_queue_tailp = &oldsd->output_queue; 12236 } 12237 /* Append NAPI poll list from offline CPU, with one exception : 12238 * process_backlog() must be called by cpu owning percpu backlog. 12239 * We properly handle process_queue & input_pkt_queue later. 12240 */ 12241 while (!list_empty(&oldsd->poll_list)) { 12242 struct napi_struct *napi = list_first_entry(&oldsd->poll_list, 12243 struct napi_struct, 12244 poll_list); 12245 12246 list_del_init(&napi->poll_list); 12247 if (napi->poll == process_backlog) 12248 napi->state &= NAPIF_STATE_THREADED; 12249 else 12250 ____napi_schedule(sd, napi); 12251 } 12252 12253 raise_softirq_irqoff(NET_TX_SOFTIRQ); 12254 local_irq_enable(); 12255 12256 if (!use_backlog_threads()) { 12257 #ifdef CONFIG_RPS 12258 remsd = oldsd->rps_ipi_list; 12259 oldsd->rps_ipi_list = NULL; 12260 #endif 12261 /* send out pending IPI's on offline CPU */ 12262 net_rps_send_ipi(remsd); 12263 } 12264 12265 /* Process offline CPU's input_pkt_queue */ 12266 while ((skb = __skb_dequeue(&oldsd->process_queue))) { 12267 netif_rx(skb); 12268 rps_input_queue_head_incr(oldsd); 12269 } 12270 while ((skb = skb_dequeue(&oldsd->input_pkt_queue))) { 12271 netif_rx(skb); 12272 rps_input_queue_head_incr(oldsd); 12273 } 12274 12275 return 0; 12276 } 12277 12278 /** 12279 * netdev_increment_features - increment feature set by one 12280 * @all: current feature set 12281 * @one: new feature set 12282 * @mask: mask feature set 12283 * 12284 * Computes a new feature set after adding a device with feature set 12285 * @one to the master device with current feature set @all. Will not 12286 * enable anything that is off in @mask. Returns the new feature set. 12287 */ 12288 netdev_features_t netdev_increment_features(netdev_features_t all, 12289 netdev_features_t one, netdev_features_t mask) 12290 { 12291 if (mask & NETIF_F_HW_CSUM) 12292 mask |= NETIF_F_CSUM_MASK; 12293 mask |= NETIF_F_VLAN_CHALLENGED; 12294 12295 all |= one & (NETIF_F_ONE_FOR_ALL | NETIF_F_CSUM_MASK) & mask; 12296 all &= one | ~NETIF_F_ALL_FOR_ALL; 12297 12298 /* If one device supports hw checksumming, set for all. */ 12299 if (all & NETIF_F_HW_CSUM) 12300 all &= ~(NETIF_F_CSUM_MASK & ~NETIF_F_HW_CSUM); 12301 12302 return all; 12303 } 12304 EXPORT_SYMBOL(netdev_increment_features); 12305 12306 static struct hlist_head * __net_init netdev_create_hash(void) 12307 { 12308 int i; 12309 struct hlist_head *hash; 12310 12311 hash = kmalloc_array(NETDEV_HASHENTRIES, sizeof(*hash), GFP_KERNEL); 12312 if (hash != NULL) 12313 for (i = 0; i < NETDEV_HASHENTRIES; i++) 12314 INIT_HLIST_HEAD(&hash[i]); 12315 12316 return hash; 12317 } 12318 12319 /* Initialize per network namespace state */ 12320 static int __net_init netdev_init(struct net *net) 12321 { 12322 BUILD_BUG_ON(GRO_HASH_BUCKETS > 12323 BITS_PER_BYTE * sizeof_field(struct gro_node, bitmask)); 12324 12325 INIT_LIST_HEAD(&net->dev_base_head); 12326 12327 net->dev_name_head = netdev_create_hash(); 12328 if (net->dev_name_head == NULL) 12329 goto err_name; 12330 12331 net->dev_index_head = netdev_create_hash(); 12332 if (net->dev_index_head == NULL) 12333 goto err_idx; 12334 12335 xa_init_flags(&net->dev_by_index, XA_FLAGS_ALLOC1); 12336 12337 RAW_INIT_NOTIFIER_HEAD(&net->netdev_chain); 12338 12339 return 0; 12340 12341 err_idx: 12342 kfree(net->dev_name_head); 12343 err_name: 12344 return -ENOMEM; 12345 } 12346 12347 /** 12348 * netdev_drivername - network driver for the device 12349 * @dev: network device 12350 * 12351 * Determine network driver for device. 12352 */ 12353 const char *netdev_drivername(const struct net_device *dev) 12354 { 12355 const struct device_driver *driver; 12356 const struct device *parent; 12357 const char *empty = ""; 12358 12359 parent = dev->dev.parent; 12360 if (!parent) 12361 return empty; 12362 12363 driver = parent->driver; 12364 if (driver && driver->name) 12365 return driver->name; 12366 return empty; 12367 } 12368 12369 static void __netdev_printk(const char *level, const struct net_device *dev, 12370 struct va_format *vaf) 12371 { 12372 if (dev && dev->dev.parent) { 12373 dev_printk_emit(level[1] - '0', 12374 dev->dev.parent, 12375 "%s %s %s%s: %pV", 12376 dev_driver_string(dev->dev.parent), 12377 dev_name(dev->dev.parent), 12378 netdev_name(dev), netdev_reg_state(dev), 12379 vaf); 12380 } else if (dev) { 12381 printk("%s%s%s: %pV", 12382 level, netdev_name(dev), netdev_reg_state(dev), vaf); 12383 } else { 12384 printk("%s(NULL net_device): %pV", level, vaf); 12385 } 12386 } 12387 12388 void netdev_printk(const char *level, const struct net_device *dev, 12389 const char *format, ...) 12390 { 12391 struct va_format vaf; 12392 va_list args; 12393 12394 va_start(args, format); 12395 12396 vaf.fmt = format; 12397 vaf.va = &args; 12398 12399 __netdev_printk(level, dev, &vaf); 12400 12401 va_end(args); 12402 } 12403 EXPORT_SYMBOL(netdev_printk); 12404 12405 #define define_netdev_printk_level(func, level) \ 12406 void func(const struct net_device *dev, const char *fmt, ...) \ 12407 { \ 12408 struct va_format vaf; \ 12409 va_list args; \ 12410 \ 12411 va_start(args, fmt); \ 12412 \ 12413 vaf.fmt = fmt; \ 12414 vaf.va = &args; \ 12415 \ 12416 __netdev_printk(level, dev, &vaf); \ 12417 \ 12418 va_end(args); \ 12419 } \ 12420 EXPORT_SYMBOL(func); 12421 12422 define_netdev_printk_level(netdev_emerg, KERN_EMERG); 12423 define_netdev_printk_level(netdev_alert, KERN_ALERT); 12424 define_netdev_printk_level(netdev_crit, KERN_CRIT); 12425 define_netdev_printk_level(netdev_err, KERN_ERR); 12426 define_netdev_printk_level(netdev_warn, KERN_WARNING); 12427 define_netdev_printk_level(netdev_notice, KERN_NOTICE); 12428 define_netdev_printk_level(netdev_info, KERN_INFO); 12429 12430 static void __net_exit netdev_exit(struct net *net) 12431 { 12432 kfree(net->dev_name_head); 12433 kfree(net->dev_index_head); 12434 xa_destroy(&net->dev_by_index); 12435 if (net != &init_net) 12436 WARN_ON_ONCE(!list_empty(&net->dev_base_head)); 12437 } 12438 12439 static struct pernet_operations __net_initdata netdev_net_ops = { 12440 .init = netdev_init, 12441 .exit = netdev_exit, 12442 }; 12443 12444 static void __net_exit default_device_exit_net(struct net *net) 12445 { 12446 struct netdev_name_node *name_node, *tmp; 12447 struct net_device *dev, *aux; 12448 /* 12449 * Push all migratable network devices back to the 12450 * initial network namespace 12451 */ 12452 ASSERT_RTNL(); 12453 for_each_netdev_safe(net, dev, aux) { 12454 int err; 12455 char fb_name[IFNAMSIZ]; 12456 12457 /* Ignore unmoveable devices (i.e. loopback) */ 12458 if (dev->netns_immutable) 12459 continue; 12460 12461 /* Leave virtual devices for the generic cleanup */ 12462 if (dev->rtnl_link_ops && !dev->rtnl_link_ops->netns_refund) 12463 continue; 12464 12465 /* Push remaining network devices to init_net */ 12466 snprintf(fb_name, IFNAMSIZ, "dev%d", dev->ifindex); 12467 if (netdev_name_in_use(&init_net, fb_name)) 12468 snprintf(fb_name, IFNAMSIZ, "dev%%d"); 12469 12470 netdev_for_each_altname_safe(dev, name_node, tmp) 12471 if (netdev_name_in_use(&init_net, name_node->name)) 12472 __netdev_name_node_alt_destroy(name_node); 12473 12474 err = dev_change_net_namespace(dev, &init_net, fb_name); 12475 if (err) { 12476 pr_emerg("%s: failed to move %s to init_net: %d\n", 12477 __func__, dev->name, err); 12478 BUG(); 12479 } 12480 } 12481 } 12482 12483 static void __net_exit default_device_exit_batch(struct list_head *net_list) 12484 { 12485 /* At exit all network devices most be removed from a network 12486 * namespace. Do this in the reverse order of registration. 12487 * Do this across as many network namespaces as possible to 12488 * improve batching efficiency. 12489 */ 12490 struct net_device *dev; 12491 struct net *net; 12492 LIST_HEAD(dev_kill_list); 12493 12494 rtnl_lock(); 12495 list_for_each_entry(net, net_list, exit_list) { 12496 default_device_exit_net(net); 12497 cond_resched(); 12498 } 12499 12500 list_for_each_entry(net, net_list, exit_list) { 12501 for_each_netdev_reverse(net, dev) { 12502 if (dev->rtnl_link_ops && dev->rtnl_link_ops->dellink) 12503 dev->rtnl_link_ops->dellink(dev, &dev_kill_list); 12504 else 12505 unregister_netdevice_queue(dev, &dev_kill_list); 12506 } 12507 } 12508 unregister_netdevice_many(&dev_kill_list); 12509 rtnl_unlock(); 12510 } 12511 12512 static struct pernet_operations __net_initdata default_device_ops = { 12513 .exit_batch = default_device_exit_batch, 12514 }; 12515 12516 static void __init net_dev_struct_check(void) 12517 { 12518 /* TX read-mostly hotpath */ 12519 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, priv_flags_fast); 12520 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, netdev_ops); 12521 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, header_ops); 12522 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, _tx); 12523 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, real_num_tx_queues); 12524 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_max_size); 12525 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_ipv4_max_size); 12526 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_max_segs); 12527 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, gso_partial_features); 12528 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, num_tc); 12529 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, mtu); 12530 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, needed_headroom); 12531 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, tc_to_txq); 12532 #ifdef CONFIG_XPS 12533 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, xps_maps); 12534 #endif 12535 #ifdef CONFIG_NETFILTER_EGRESS 12536 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, nf_hooks_egress); 12537 #endif 12538 #ifdef CONFIG_NET_XGRESS 12539 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_tx, tcx_egress); 12540 #endif 12541 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_tx, 160); 12542 12543 /* TXRX read-mostly hotpath */ 12544 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, lstats); 12545 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, state); 12546 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, flags); 12547 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, hard_header_len); 12548 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, features); 12549 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_txrx, ip6_ptr); 12550 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_txrx, 46); 12551 12552 /* RX read-mostly hotpath */ 12553 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, ptype_specific); 12554 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, ifindex); 12555 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, real_num_rx_queues); 12556 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, _rx); 12557 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_max_size); 12558 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, gro_ipv4_max_size); 12559 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, rx_handler); 12560 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, rx_handler_data); 12561 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, nd_net); 12562 #ifdef CONFIG_NETPOLL 12563 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, npinfo); 12564 #endif 12565 #ifdef CONFIG_NET_XGRESS 12566 CACHELINE_ASSERT_GROUP_MEMBER(struct net_device, net_device_read_rx, tcx_ingress); 12567 #endif 12568 CACHELINE_ASSERT_GROUP_SIZE(struct net_device, net_device_read_rx, 92); 12569 } 12570 12571 /* 12572 * Initialize the DEV module. At boot time this walks the device list and 12573 * unhooks any devices that fail to initialise (normally hardware not 12574 * present) and leaves us with a valid list of present and active devices. 12575 * 12576 */ 12577 12578 /* We allocate 256 pages for each CPU if PAGE_SHIFT is 12 */ 12579 #define SYSTEM_PERCPU_PAGE_POOL_SIZE ((1 << 20) / PAGE_SIZE) 12580 12581 static int net_page_pool_create(int cpuid) 12582 { 12583 #if IS_ENABLED(CONFIG_PAGE_POOL) 12584 struct page_pool_params page_pool_params = { 12585 .pool_size = SYSTEM_PERCPU_PAGE_POOL_SIZE, 12586 .flags = PP_FLAG_SYSTEM_POOL, 12587 .nid = cpu_to_mem(cpuid), 12588 }; 12589 struct page_pool *pp_ptr; 12590 int err; 12591 12592 pp_ptr = page_pool_create_percpu(&page_pool_params, cpuid); 12593 if (IS_ERR(pp_ptr)) 12594 return -ENOMEM; 12595 12596 err = xdp_reg_page_pool(pp_ptr); 12597 if (err) { 12598 page_pool_destroy(pp_ptr); 12599 return err; 12600 } 12601 12602 per_cpu(system_page_pool, cpuid) = pp_ptr; 12603 #endif 12604 return 0; 12605 } 12606 12607 static int backlog_napi_should_run(unsigned int cpu) 12608 { 12609 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 12610 struct napi_struct *napi = &sd->backlog; 12611 12612 return test_bit(NAPI_STATE_SCHED_THREADED, &napi->state); 12613 } 12614 12615 static void run_backlog_napi(unsigned int cpu) 12616 { 12617 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 12618 12619 napi_threaded_poll_loop(&sd->backlog); 12620 } 12621 12622 static void backlog_napi_setup(unsigned int cpu) 12623 { 12624 struct softnet_data *sd = per_cpu_ptr(&softnet_data, cpu); 12625 struct napi_struct *napi = &sd->backlog; 12626 12627 napi->thread = this_cpu_read(backlog_napi); 12628 set_bit(NAPI_STATE_THREADED, &napi->state); 12629 } 12630 12631 static struct smp_hotplug_thread backlog_threads = { 12632 .store = &backlog_napi, 12633 .thread_should_run = backlog_napi_should_run, 12634 .thread_fn = run_backlog_napi, 12635 .thread_comm = "backlog_napi/%u", 12636 .setup = backlog_napi_setup, 12637 }; 12638 12639 /* 12640 * This is called single threaded during boot, so no need 12641 * to take the rtnl semaphore. 12642 */ 12643 static int __init net_dev_init(void) 12644 { 12645 int i, rc = -ENOMEM; 12646 12647 BUG_ON(!dev_boot_phase); 12648 12649 net_dev_struct_check(); 12650 12651 if (dev_proc_init()) 12652 goto out; 12653 12654 if (netdev_kobject_init()) 12655 goto out; 12656 12657 for (i = 0; i < PTYPE_HASH_SIZE; i++) 12658 INIT_LIST_HEAD(&ptype_base[i]); 12659 12660 if (register_pernet_subsys(&netdev_net_ops)) 12661 goto out; 12662 12663 /* 12664 * Initialise the packet receive queues. 12665 */ 12666 12667 flush_backlogs_fallback = flush_backlogs_alloc(); 12668 if (!flush_backlogs_fallback) 12669 goto out; 12670 12671 for_each_possible_cpu(i) { 12672 struct softnet_data *sd = &per_cpu(softnet_data, i); 12673 12674 skb_queue_head_init(&sd->input_pkt_queue); 12675 skb_queue_head_init(&sd->process_queue); 12676 #ifdef CONFIG_XFRM_OFFLOAD 12677 skb_queue_head_init(&sd->xfrm_backlog); 12678 #endif 12679 INIT_LIST_HEAD(&sd->poll_list); 12680 sd->output_queue_tailp = &sd->output_queue; 12681 #ifdef CONFIG_RPS 12682 INIT_CSD(&sd->csd, rps_trigger_softirq, sd); 12683 sd->cpu = i; 12684 #endif 12685 INIT_CSD(&sd->defer_csd, trigger_rx_softirq, sd); 12686 spin_lock_init(&sd->defer_lock); 12687 12688 gro_init(&sd->backlog.gro); 12689 sd->backlog.poll = process_backlog; 12690 sd->backlog.weight = weight_p; 12691 INIT_LIST_HEAD(&sd->backlog.poll_list); 12692 12693 if (net_page_pool_create(i)) 12694 goto out; 12695 } 12696 if (use_backlog_threads()) 12697 smpboot_register_percpu_thread(&backlog_threads); 12698 12699 dev_boot_phase = 0; 12700 12701 /* The loopback device is special if any other network devices 12702 * is present in a network namespace the loopback device must 12703 * be present. Since we now dynamically allocate and free the 12704 * loopback device ensure this invariant is maintained by 12705 * keeping the loopback device as the first device on the 12706 * list of network devices. Ensuring the loopback devices 12707 * is the first device that appears and the last network device 12708 * that disappears. 12709 */ 12710 if (register_pernet_device(&loopback_net_ops)) 12711 goto out; 12712 12713 if (register_pernet_device(&default_device_ops)) 12714 goto out; 12715 12716 open_softirq(NET_TX_SOFTIRQ, net_tx_action); 12717 open_softirq(NET_RX_SOFTIRQ, net_rx_action); 12718 12719 rc = cpuhp_setup_state_nocalls(CPUHP_NET_DEV_DEAD, "net/dev:dead", 12720 NULL, dev_cpu_dead); 12721 WARN_ON(rc < 0); 12722 rc = 0; 12723 12724 /* avoid static key IPIs to isolated CPUs */ 12725 if (housekeeping_enabled(HK_TYPE_MISC)) 12726 net_enable_timestamp(); 12727 out: 12728 if (rc < 0) { 12729 for_each_possible_cpu(i) { 12730 struct page_pool *pp_ptr; 12731 12732 pp_ptr = per_cpu(system_page_pool, i); 12733 if (!pp_ptr) 12734 continue; 12735 12736 xdp_unreg_page_pool(pp_ptr); 12737 page_pool_destroy(pp_ptr); 12738 per_cpu(system_page_pool, i) = NULL; 12739 } 12740 } 12741 12742 return rc; 12743 } 12744 12745 subsys_initcall(net_dev_init); 12746